A Comprehensive Survey of Machine Learning Applied to Radar Signal Processing
Ping Lang, Xiongjun Fu, Marco Martorella, Jian Dong, Rui Qin, Xianpeng Meng, Min Xie
aa r X i v : . [ ee ss . SP ] S e p JOURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 1
A Comprehensive Survey of Machine LearningApplied to Radar Signal Processing
Ping Lang,
Student Member, IEEE , Xiongjun Fu,
Member, IEEE ,Marco Martorella,
Fellow, IEEE , Jian Dong, Rui Qin, Xianpeng Meng, and Min Xie
Abstract —Modern radar systems have high requirements interms of accuracy, robustness and real-time capability whenoperating on increasingly complex electromagnetic environments.Traditional radar signal processing (RSP) methods have shownsome limitations when meeting such requirements, particularlyin matters of target classification. With the rapid developmentof machine learning (ML), especially deep learning, radar re-searchers have started integrating these new methods whensolving RSP-related problems. This paper aims at helping re-searchers and practitioners to better understand the applicationof ML techniques to RSP-related problems by providing acomprehensive, structured and reasoned literature overview ofML-based RSP techniques. This work is amply introduced byproviding general elements of ML-based RSP and by stating themotivations behind them. The main applications of ML-basedRSP are then analysed and structured based on the applicationfield. This paper then concludes with a series of open questionsand proposed research directions, in order to indicate currentgaps and potential future solutions and trends.
Index Terms —Radar signals classification and recognition,SAR/ISAR images processing, radar anti-jamming, machinelearning, deep learning
I. I
NTRODUCTION R ADAR offers special advantages with respect to othertypes of sensors including all-day, all-weather operations,long detection distance and, depending on the frequency used,penetration. Moreover, radar can often be carried by a numberof platforms, spanning from classic naval and airborne to morerecent space-borne, UAVs, such as drones, and high-altitudeplatforms (HAPs). The ensemble of these characteristics canbe exploited for military scenarios, such as target detection,tracking and recognition, and for civil scenarios, such asland use and classification, disaster assessment, urban andnon-urban monitoring, making radar the perfect sensor fordual use applications [1], [2]. Radar signal processing (RSP)is one of the key aspects that characterize the radar field[3] as its development allows for radar performances to bemaximised and for several capabilities to be enabled, includingthe ability to operate in spectrally congested and contestedscenarios and complex and dynamically changing environment[4], [106]. Artificial Intelligence (AI) has pushed the researchand development in many fields [5], including, among others,speech signal processing (SSP), computer vision (CV) andnatural language processing (NLP). Such domains include
This work was supported by the National Natural Science Foundation ofChina under Grant 61571043 and 111 Project of China under Grant B14010. (Corresponding author: Xiongjun Fu)
P. Lang, X. Fu are with School of Information and Electronics, BeijingInstitute of Technology, Beijing 100081, China (E-mails: { langping911220,fuxiongjun, } @bit.edu.cn). logic programming, expert system, pattern recognition, ma-chine learning (ML) and reinforcement learning [6]. Machinelearning (ML), and especially deep learning (DL) [7], [8],has achieved great breakthroughs thanks to large investmentsfrom a number of countries and through a pervasive coop-eration of the scientific community. More specifically, ML-based RSP has been targeted by many to attempt to improvetraditional RSP solutions and overcome their limitations. As ademonstration of the interest in this field, in the recent years,the Defense Advanced Research Projects Agency (DARPA)has launched many projects in this field, such as the radiofrequency machine learning system (RFMLS) project [9]–[12],the behavior learning for adaptive electronic warfare (BLADE)project [13], [14], and the adaptive radar countermeasures(ARC) project [15]. In addition to DARPA’s projects, there isample support from the scientific literature, such as radar emit-ter recognition and classification [110], [147], [150], [152],radar image processing (e.g., synthetic aperture radar (SAR)image denoising [273]–[276], [279], data augmentation [251]–[255], automatic target recognition (ATR) [304], [310]–[316],[326], target detection [585], [587], also with specific emphasison ship detection [472]–[474], [476], [477], anti-jamming[576], optimal waveform design [580], array antenna selection[586], and cognitive electronic warfare (CEW) [584]. TheseML algorithms include traditional machine learning (e.g.,support vector machines (SVMs), decision tree (DT), randomforest (RF), boosting methods), and deep learning (e.g., deepbelief networks (DBNs), autoencoders (AEs), convolutionalneural networks (CNNs), recurrent neural networks (RNNs),generative adversarial networks (GANs)). This survey paperhas comprehensively reviewed state-of-the-art of ML-basedRSP algorithms, including traditional ML and DL. A. Motivation
Due to the large success of ML in many domains, theradar community has started applying ML-based algorithmsto classic and new radar research domains to tackle traditionaland new challenges from a novel prospective. Being ML arelatively new paradigm, the research results that have beenobtained have not been systematically surveyed and analyzed.A thorough and reasoned review of new technologies is keyfor providingi) a solid basis for new researchers and practitioners whoare approaching this field for the first time;ii) an important reference for more experienced researcherswho are working in this field;iii) existing terms for comparison for newly developed ML-based algorithms;iv) means to identify gaps;
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 2 v) a full understanding of strengths and limitations of ML-based approaches.
B. Related works
This section will briefly survey some this topic-relatedreview scientific literatures. i) ML algorithms and applications
There are many reviewpapers either about the development of ML algorithms, suchas DL [8], [17], [19], deep reinforcement learning (DRL)[16], transfer learning (TL) [18], GANs [20], developmentsof CNNs [38], efficient processing technologies of DNN[39], adversarial learning for DNN classification [40], neuralnetworks model compression and hardware acceleration [41]or applications in special topics such as ML applied to medicalimage processing [21], [22], robotics [26], agriculture [23],sentiment analysis [24], object detection [42], [43].As a most popular DNN model, CNN has been successfullyapplied in most of ML tasks. In [38], the authors compre-hensively investigated the state-of-the-art technologies aboutthe development of CNN. This paper systematically intro-duced the CNN models from LeNet to latest networks suchas GhostNet, including one-dimension (1D), two-dimension(2D), and multi-dimension (multi-D) convolutional modelsand their applications, such as 1D, 2D and multi-D modelscan be applied in time series prediction and signal iden-tification, image processing, and human action recognition,X-ray, computation tomography (CT), respectively. Besides,some prospective trends have been proposed such as modelcompression [41], security, network architecture search [594],and capsule neural network [25].TL aims to solve insufficient training data problem, whichalso used in RSP domain, such as radar emitter recognition[206], micro-doppler for motion classification [542], [544],SAR image processing with limited labeled data [305], [306],[329]. A TL-related review was developed in [18], whichcategorized the TL techniques as four classes: instances-based, mapping-based, network-based, and adversarial-based,respectively.Object detection, as one of most important tasks of CV,is a fundamental and challengeable task, which not onlyconcentrates on classifying different images but also triesto precisely estimate the concepts and locations of objectscontained in each image [42]. The authors in [42], [43] havestudied the latest development of object detection in the pastfew years. These review papers have covered many aspects ofobject detection, including detection frameworks (such as R-CNN, Fast R-CNN, Faster R-CNN, Mask R-CNN, YOLOv1-v4), training strategy, evaluation metrics, and the analysis ofsome typical object detection examples, such as salient objectdetection, face detection and pedestrian detection. ii) Remote sensing
Besides, some review papers, focusedon ML applied to remote sensing (RS) domain, have beenpublished in [27]–[31]. These survey papers investigated thestate-of-the-art technologies of ML to solve the challenges inRS domain, such as RS image processing (e.g., hyperspectralimage, SAR image, hyper resolution satellite image, 3D re-construction), target recognition, scene understanding, objectdetection and segmentation. The challenges of using DL for RS data analysis wereanalyzed in [27], and then the recent advances in imagesclassification, recognition, detection, multi-model data fusion,and 3D reconstruction were reviewed . The DL models mainlyincluded AEs and CNNs. The authors in [28] surveyed therecent developments of RS field with DL and provided atechnique tutorial on the design of DL-based methods forprocessing the optical RS data, including image preprocess-ing, pixel-based classification, target recognition, and sceneunderstanding. The comprehensive survey of state-of-the-artDL in RS research was developed in [29], which focusedon theories, tools, challenges for the RS community, andspecifically discussed unsolved challenges and opportunities,such as inadequate data sets, human understandable solutionsfor modeling physical phenomena. In [31], the authors sys-tematically reviewed the DL in RS applications by meta-analysis method containing image fusion, registration, sceneclassification and object detection, semantic segmentation,and even accuracy assessment. The recent progress of RSimage scene classification, especially DL-based methods wassurveyed in [32]. In addition, a large-scale remote sensingimage scene classification (RESISC) benchmark data set,termed “NWPU-RESISC45” was proposed. The traditionalML algorithms applied to classification of RS research wasalso investigated in [30], including SVM, boosted DTs, RF,artificial neural network (ANN), K nearest neighbor (K-NN).The study aspects contained the selection of classifier, therequirements of training data, definition of parameters, featurespace operation, model interpretability, and computation costs.Some key findings such as SVM, RF, and boosted DTs havehigher accuracy for classification of remotely sensed data,compared to alternative machine classifiers such as a singleDT and K-NN.A comprehensive state-of-the-art survey for SAR-ATR tech-niques was developed in [34], which was categorized to model-based, semi-model-based, and feature-based. These SAR-ATRtechniques, however, were unilaterally based on pattern recog-nition or prior knowledge. The AE model and its variantsapplied to RS and SAR images interpretation was investigatedin [33], including original AE, sparse AE, denoising AE, con-volutional AE, variational AE, and contrastive AE. The authorsin [35] surveyed temporal developments of optical satellitecharacteristics and connected these with vessel detection andclassification after analyzed 119 selected literatures. Althoughthere are some review papers about RS domain based on MLalgorithms, as a subset of RS, the comprehensive survey ofML algorithms applied to RSP has not emerged so far. iii) Multi-representation learning algorithms
There arealso some other survey papers related the topic of this area,such as multi-view learning (MVL) [36], multi-task learning(MTL) [37].MVL and MTL have rapidly grown in ML and datamining in the past few years, which can obviously improveperformance of model learning. In RSP domain, these relatedmethods are popular in DL-based SAR-ATR, e.g., [325],[328], [331], [336], [337], [340], [343], [347]. Therefore, itis necessary to make a brief introduction about the reviewpapers in MVL [36] and MTL [37]. MVL is concerned as the
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 3 !" !" !!" 4..5 *.%$)!)/ !) 001206 !" %%9 !" 4%’%7.’7 %)& %8/+.)’%’!()3 !!" 190 !+%/.7 &.)(!7!)/3 !!!" 190:9 %%; !" <190 !+%/!)/3 !!" <190 ’%$/.’7 &.’.,’!() %)& $.,(/)!’!() %%< % !" =%++!)/ ($ !)’.$>.$.),. ,*%77!>!,%’!() %)& $.,(/)!’!()3!!" 9)’!:?%++!)/ %)& !)’.$>.$.),. +!’!/%’!() 7’$%’./!.76 !" !" @%;.>($+ (5’!+%* &.7!/)3 !!" 2(/)!’!;. .*.,’$()!, A%$>%$.3!!!" 9)’.))% %$$%B &.7!/)3!;" 0%&%$ 75.,’$8+ %**(,%’!()3 ;" C(;!)/ ’%$/.’ !)&!,%’!()DC !" E)&:’(:.)& 8)!>!.& !)’.**!/.)’ &.’.,’!() %$,-!’.,’8$.3!!" !!!" E)&:’(:.)& >$%+.A($F (> 8)!>!.& &.’.,’!()G ’$%,F!)/ %)& $.,(/)!’!()6 ; !" E%7!)/ *!+!’%’!() (> &%’%7.’73 !!" !;" 2(/)!’!;. A%;.>($+ &.7!/)3;" <)’.**!/.)’ %)’!:?%++!)/6 Fig. 1. The overview contents of this paper. problem of learning representations (or features) of the multi-view data that facilitates extracting readily useful informationwhen developing prediction models [36].According to the state-of-the-art overview on MVL studiedin [36], multi-view representation alignment and multi-viewrepresentation fusion were two categories of MVL. The formeraims to capture the relationships among multiple differentviews through feature alignment, including multi-modal topiclearning, multi-view sparse coding, and multi-view latentspace Markov networks. The latter seeks to fuse the separatefeatures, learned from multiple different views, into a singlecompact representation, including multi-modal AEs, multi-view CNNs, and multi-modal RNNs.MTL can be roughly defined as: a some task learningcan improve generalization capability by shared representationbetween related tasks or optimize multi-loss function simul-taneously. A comprehensive survey on MTL in deep neuralnetworks (DNNs) was developed in [37], which introduced(i) two common MTL methods in DL, i.e., hard and softparameters sharing, (ii) MTL neural network models and non-NN models, such as block-sparse regularization, learning tasksrelationship, and (iii) auxiliary tasks in order to reap thebenefits of multi-task learning.
C. Contributions and Organization
Motivated by the research community and our research in-terests, this article collects state-of-the-art achievements aboutML-based RSP algorithms from public databases such as
IEEEXplore, Web of Science, and dblp , most of which come fromrecent 5 years, i.e., from 2015 to 2020. We systematicallyanalyze these findings on this research domain, to pave theaccess to promising and suitable directions for future research.Hopefully, this paper can help relative researchers and practi-tioners to quickly and effectively determine potential facts ofthe this topic by clearly knowing about key aspects and relatedbody of research.In this consideration, we make mainly three contributions:(i) Based on a deep literatures analysis of more than 600papers, we firstly provide an systematical overview of theexisting approaches of ML-based RSP domain from differentperspectives; (ii) We propose a comprehensive background regardingthe main concepts, motivations, and implications of enablingintelligent algorithm in RSP;(iii) A profound discussion about the future promisingresearch opportunities and potential trends in this field isproposed.Accordingly, the reminder of this review article is organizedas follows. Section II briefly introduces the basic principles oftypical ML algorithms; section III surveys the latest develop-ments in radar radiation sources classification and recognition;section IV investigates state-of-the-art achievements in radarimage processing; section V investigates the developments ofanti-jamming and interference mitigation; other RSP-relatedresearch that does not fall in previous categories, such aswaveform design, anti-interference, has been reviewed insection VI; section VII profoundly discusses open problemsand possible promising research directions, in order to indicatecurrent gaps and potential future solutions and trends. Finally,the conclusion of this article is drawn in section VIII. Theoverview contents of this paper is shown in Fig. 1.II. T HE B ASIC P RINCIPLES OF T YPICAL M ACHINE L EARNING A LGORITHMS
ML has achieved great success in many domains, mainlyrelated to three determining factors: data, model algorithm, andcomputation power. As a data-driven pattern, big data is thebasic motivation for development of ML. Computation poweris supported by hardware equipments to drive ML modeltraining, such as graphical processing units (GPUs), tensorprocessing units (TPUs), Kunpeng 920 produced by Huaweicorporation. This section will briefly introduce the RSP-relatedtypical ML model algorithms.
A. Traditional Machine Learning Models
1) Support Vector Machines (SVMs) .Support Vector Machines are the most popular ML algo-rithms for binary classification [44], especially high-efficientlyin solving non-linear binary classification issues, through theprojection of low dimensional feature space to a higher onewith kernel function [45] (e.g., polynomial kernels, radial basisfunction (RBF) kernel, Gaussian kernel). SVMs address theclassification problem by finding an optimal hyperplane in
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 4 the feature space to maximize the samples margin betweenthe support vectors of two classes, as shown in Fig. 2. Theoptimal problem can be expressed as in Eq.(1), which is aconvex optimization problem, and the sequence minimizationoptimization (SMO) [47] can be used as an optimizationalgorithm. SVMs have been widely applied to radar emitterclassification and recognition [170], [192]–[194]. hyperplane
Support vectors class1 class2wx+b
Fig. 2. The diagram of SVM, d = d . min 12 k ω k s.t. y ( i ) ( (cid:10) ω , x i (cid:11) + b ) ≥ i = 1 , ...m ) , (1)where ω and b are the hyperplane parameters, m is the numberof samples, x , y are the samples and the labels, respectively.When the classes do not have an explicit classification hyper-plane, i.e., inherently not separable. Soft-SVMs can be used totackle this issue. This means that a small number of samplesis allowed to fall into the wrong side. The objective of soft-SVMs adds a penalty term based on SVMs to restrict the slackterm ε , as follow: min ( 12 k ω k + C X ε i ) s.t. y ( i ) ( (cid:10) ω , x i (cid:11) + b ) ≥ − ε i ( i = 1 , ...m ) ε i ≥ max n , − y ( i ) ( (cid:10) ω , x i (cid:11) + b ) o , (2)where C is the penalty term.
2) Decision Trees (DTs) .Decision Trees are intuitively the simplest case of MLalgorithm. They are suitable for addressing these situationswhere the labels of data are non-continuous. DTs adopt if-then rules to split the input data according to features andsuitable threshold values based on a binary tree structure [48].The root nodes, middle nodes, and leaf nodes represent inputdata, features and threshold attributes and outputs, respectively.Each branch represents an output of the discrimination process.The loss function is usually implemented as a mean squareerror (MSE) for regression and cross entropy (CE) for classifi-cation. DTs typically use a limitation of the tree structure depth and pruning operations to address the overfitting problem.Although pruning will reduce the task accuracy to someextent, it generally improves the generalization. Informationentropy-based ID3 [48], C4.5 [49], and Gini coefficient-basedclassification and regression tree (CART) [50] are usuallythe optimization algorithms that are implemented during thetraining process. DTs has been applied to radar emitter clas-sification and recognition [203].
3) Boosting Ensemble Learning .The Ensemble Learning (EL) [51] builds multi-classifier tojointly make prediction of inputs. The advantages of ensemblelearning are as follow: (i) improving prediction accuracy withjoint decision; (ii) can easily deal with either large or smalldatasets, i.e., large dataset can be divided into multiple subsetsto build a multi-classifier, small dataset can be sampled toreform multiple datasets to establish a multi-classifier; and (iii)suitable to address the complex decision boundary problems,homologous and heterogeneous datasets. EL can be catego-rized into two classes: bootstrap (such as random forest) and boosting (such as adaboost [53], gradient boosting decisiontree (GBDT) [55], extreme GBDT (XGBoost) [54]). Gradientboosting methods [55], [207] were used as classification modelin radar emitter recognition.
Random forest (RF)
As one of the ensemble learning algo-rithms [52], RF is a bootstrap ensemble classifier, consistingof relatively independent multi-CART, to overcome the highprediction error problem with a single DT. Every sub-DT isa weak learning model as a part of an over learning task,trained by a random subset bootstrapped from the trainingdataset, and determined splitpoint with random features. Thefinal prediction output is determined by voting rules with allDTs. RF may reach the global optimum, instead of a localoptimum as in the case of a single tree. The radar signalsrecognition based on RF models was proposed in [170] toobtain comparable performance.
Adaboost
Adaboost, i.e., adaptive boosting, which firstlyproduce a set of hypothesis functions by repeatedly using basiclearning algorithm based on multi-sampled training data. Then,these hypothesis functions are connected to ultimately forman ensemble learner via linear weighted vote rules [53]. AnAdaBoost algorithm was employed as a classifier in [209] tocomplete the different types recognition of radar signals with1D harmonic amplitude data sets.Given the hypothesis function H = { h ( x ) : x → R } andunknown data x , h ( x ) donates weak learners or base learners,then the ultimate ensemble learner can be given by: F ( x ) = T X t =1 α t h t ( x ) , (3)where α t is the connection coefficients of t -th iteration, T is the number of iteration. α = [ α , α , .., α T ] and h ( x ) areoptimally generated during the minimization of loss function C , as showed in Eq.(4). Initially, the weight of every sampleis set equal to N , N being the number of samples. Whenthe sample is misclassified, it gets a larger weight in thefollowing iterations, the base learner is forced to focus on OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 5 these hard-to-classify cases in the subsequent training steps.This characterizes the adaptation of boosting methods. C = 1 N N X n =1 exp( − y n F ( x n )) , (4)where y n ⊂ { +1 , − } is the label of data x n . GBDT
As a residual learning type, the prediction scoreof GBDT [55] is determined by summing up the scoresof a multi-CART regression tree, instead of a classifica-tion tree. In detail, adding a new tree structure to learnthe residual (i.e., the gap between the prediction and theactual value) of previous tree at each iteration based onnegative gradient learning, to iteratively approach the actualvalue. For a given dataset with n samples and m features D = { ( x i , y i ) } ( i = 1 ...n, x i ∈ R m , y i ∈ R ) which uses K additive tree functions to predict the output (take a regressiontree as an example) [54]: ˆ y i = φ ( x i ) = K X k =1 f k ( x i ) , f k ∈ Γ , (5)where Γ = { f ( x ) = ω q ( x ) } ( q : R m → T, ω ∈ R T ) is thespace of regression tree. q represents the structure of eachtree that maps an example to the corresponding leaf index. T is the number of leave nodes in the tree. Each f k correspondsto an independent tree structure q and leaf weights ω . Let y i and ˆ y i be the actual and prediction values, then we minimizethe following loss function: L ( φ ) = X i l ( y i , ˆ y i ) ( i = 1 ...n, y i ∈ R ) , (6)where l is a loss function, usually the MSE. XGBoost
As an implementation of gradient tree boosting,XGBoost [54], an end-to-end scalable tree boosting system, iswidely used in data mining. As one of the most popular MLmodel, it provides the state-of-the-art performance in manyKaggle competitions in recent years. For example, 17 solutionsused XGBoost (eight solely used XGBoost, while otherscombined XGBoost with neural networks) among the 29 chal-lengeable winning solutions at 2015 Kaggle competition [54].XGBoost was also used in the top-10 in the KDDCup 2015by each award-winning team [54]. In addition, the authors in[180] used weighted-XGBoost for Radar emitter classification.XGBoost’s widespread scalability as its one of the mostimportant factor of success, which can scale to billions ofexamples in distributed or memory-limited setting and havehigher computation efficiency than existing popular solutionson a single machine. Compared to GBDT, XGBoost adds apenalty term (i.e., regularized term) in objective function toovercome overfitting, and introduces the first and second ordergradient in objective based on Taylor expansion. The XGBoostminimizes the following objective, L ( φ ) = X i l ( y i , ˆ y i ) + X k Ω( f k ) , Ω( f ) = γT + 12 λ k ω k , (7) Inputs Hidden
Outputs
Sum f w w w w n ba a a a n nnnnnn Output a b
Fig. 3. The diagram of ANN,(a)single neutron,(b)artificial neural networkwith one-hidden layer. where Ω is the regularized term to penalize the complexityof the model, usually l norm or l norm . The optimizationalgorithm is residual learning iteratively between the adjacentsub-model, and let ˆ y t − i be as the prediction of i -th instanceat ( t − -th iteration, we minimize the following objective, L ( t ) = n X i =1 l ( y i , ˆ y t − i + f t ( x i )) + Ω( f t ) , (8)where f t represents the residual between ( t − -th and t -thiterations. Inspired with Taylor expansion (the second order ex-pansion): f ( x + ∆ x ) ≈ f ( x ) + f ′ ( x )∆ x + f ′′ ( x )∆ x . Theabove equation can be rewritten as follow, L ( t ) ≈ n X i =1 [ l ( y i , ˆ y t − i ) + g i f t ( x i ) + 12 h i f t ( x i )]+Ω( f t ) , (9)where g i = ∂ ˆ y t − l ( y i , ˆ y t − i ) and h i = ∂ y t − i l ( y i , ˆ y t − i ) are thefirst and second order gradient statistics on the loss function,respectively. Removing the constant term [ l ( y i , ˆ y t − i )] to sim-plify the Eq.(7) by L ( t ) = n X i =1 [ g i f t ( x i ) + 12 h i f t ( x i )] + Ω( f t ) . (10)Define C = { j | q ( x i ) = j } as the set of leaf nodes j , Eq.(8)can be rewritten as L ( t ) = n X i =1 [ g i f t ( x i ) + 12 h i f t ( x i )] + γT + 12 λ T X j =1 ω j = T X j =1 [( X i ∈ C g i ) ω j + 12 ( X i ∈ C h i ) ω j ] + γT + 12 λ T X j =1 ω j = T X j =1 [( X i ∈ C g i ) ω j + 12 ( X i ∈ C h i + λ ) ω j ] + γT . (11)When fixing a tree q ( x ) , the optimal score ω ∗ j of leaf nodes j is given by ω ∗ j = − P i ∈ C g i P i ∈ C h i + λ . (12) OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 6
The optimal objective at t -th iteration is given by L t ( q ) = − T X j =1 ( P i ∈ C g i ) P i ∈ C h i + λ + γT . (13)Eq.(10) can be used as a scoring function to measure thequality of a tree structure q . A greedy algorithm is used tosearch for optimal tree structure C L and C R ( C = C L ∪ C R )are sets of left and right nodes after being split. The reductionof loss after being split is given by L split = 12 [ ( P i ∈ C R g i ) P i ∈ C R h i + λ + ( P i ∈ C L g i ) P i ∈ C L h i + λ − ( P i ∈ C g i ) P i ∈ C h i + λ ] − γ. (14)
4) Artificial Neural Networks (ANNs) .Inspired by human’s brain neural network, ANN, a simpli-fied mathematical analogue of human’s neural network [56], isused to process the input information by the layer-wise stylefor regression and classification tasks, as shown in Fig. 3. Abasic ANN has one input layer, one hidden layer, and outputlayer, each of which has many artificial neutrons. The numberof neurons is determined by dimension of input data for inputlayer, number of classes for output layer, and alternative forhidden layer. All neurons in one layer are connected to allneurons in all adjacent layers (i.e., fully connection) withweights, bias, and non-linear activation function for everyneuron. Obviously, more neurons in hidden layer or morehidden layers, will rapidly increase the ability of informationprocessing because of improved power of feature extraction ofdata, which characterizes the deep learning algorithm (will beintroduced in next subsection). The optimal training method isbackpropagation algorithm to iteratively update the parametersof ANN. ANNs were widely used in radar emitter recognition[197]–[201]. ??????
Fig. 4. The diagram of K-NN.
5) K-Nearest Neighbors (K-NNs) .As an instance-based learning style, K-NNs are not likeother classifiers to explicitly train a classification model toclassify unknown samples [57]. Instead, samples have afore-hand classes and features, and the class of unknown sample is determined by K nearest neighbors surrounding it, which isevaluated by the distances between feature spaces of aforehandsamples and unknown samples (such as Euclidean distance,cosine distance, correlation, Manhattan distance). The un-known sample belongs to the class where the highest frequencyin K nearest neighbor samples. For example, in Fig. 4, whenK = 3, the color of yellow cycle classified to red, and the colorof yellow cycle is classified to blue when K = 7. K is verysignificant for classification, which usually starts with K = 1,iteratively finding the smallest error with increment 1. K-NNis adopted to classify instantaneous transient signals based onradio frequency fingerprint extraction in [181].
B. Deep Learning Models
DL models, also called DNN, consist of multi-layer ANN,i.e., input layers, multi-hidden layer, and output layer, whichtransform input data (e.g., images, sequences) to outputs (e.g.,classes) with the high-level feature representation learning bymulti-hidden layers.In 2006, Hinton has successfully achieved training of DBNwith gradient decent backpropagation algorithm, and experi-ments results determined promissing performance in CV tasks[58]. This breakthrough quickly draw insights from the indus-trial and academics. Especially, CNN-based AlexNet architec-ture has firstly won the human in the competition of ImageNetcontest in 2012 [59]. DL has developed rapidly in manydomains, such as speech recognition [60], image processing[59], [61]–[63], audio signal processing [64], [65], videoprocessing [66], [67], and NLP [68]–[70]. In the followingyears, many novel DL architectures and domain achievementshave developed, including CNNs, RNNs, and GANs.The remainder of this section is contributed to briefly in-troduce several commonly used DL models in RSP, includingunsupervised AEs, DBNs and GANs, and supervised CNNs,RNNs.
1) Restricted Boltzmann Machines (RBMs) and DeepBelief Networks (DBNs) .A restricted Boltzmann machine (RBM), composed by a vis-ible layer x and a hidden layer h , and symmetric connectionsbetween these two layers represented by a weight matrix W ,is a generative stochastic undirectional neural network [64].The joint distribution of visible and hidden units is defined byits energy function as follow [71], [72] P ( v, h ) = 1 Z e ( − E ( v,h ) ) , (15)where Z is the partition function. If the visible units are binary-value, E ( v, h ) can be defined as E ( v, h ) = − X i,j v i W ij h j − X j b j h j − X i c i v i , (16)where b j and c i are hidden unit bias and visible unit biasrespectively. b, c, W are the parameters of RBM model.A DBN can be viewed as a stacked structure of multi-RBM model [58], [64], [73], which is regarded as a generativeprobabilistic graphical model. DBN can break the limitationof RBM representation with a fast training algorithm [58].The RBM and DBN examples are shown in Fig. 5. In RSP OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 7 v hWv h h h h RBM1
RBM2 RBM3 ab Fig. 5. The RBM and DBN model, (a) RBM,(b) DBN. domain, DBN has been used to radar emitter recognitionand classification [234]–[236], HRRP-ATR [517], [518], SAR-ATR [312].
2) Autoencoders (AEs) .AEs are basically unsupervised learning algorithms, whichnormally accomplish the tasks of data compression and di-mensionality reduction in unsupervised manner. An AE modelconsists of three opponents: encoder, activation function, anddecoder, as shown in Fig. 6 and Fig. 7 [33]. !" (cid:165) % !" (cid:169) % ’ (" (cid:262) &)*+,-./ 0+123412,* 5.+,-./6*7819 %!& ( :817819 %!;&" <,99 Fig. 6. The general autoencoder model.
Encoder f can be regarded as a linear feed-forward filterof input x determined by weight matrix W and bias b , i.e., f = W x + b .Activation function σ performs a non-linear mapping thattransforms the f into latent representation h of input x at therange of [0 , , i.e., h = σ ( W x + b ) .The decoder g is a reverse linear filter to produce thereconstruction e x of the input x , i.e., e x = g ( W T h + b ′ ) A loss function L is used to measure how close the AE canreconstruct the output e x , i.e., L ( e x, x ) . The training processingis minimizing the loss between e x and x , i.e., min ( L ( e x, x )) . Sparse autoencoder (SAE)
In order to accelerate the train-ing of AE model, SAE characterizes by adding sparsity con-straints to the hidden layers, and only activating the neurons
Encoder Decoder
Fig. 7. The fully connection neural network model of Autoencoder. whose outputs are close to 1. Therefore, the only small amountof parameters was needed to learn greatly reduce training time.
Denoising autoencoder (DeAE)
To increase the robustnessof AE with small various input data, DeAE has been proposedin [74]. Before entering into the input layer, the original input x is corrupted as x ′ . Binary noisy and Gaussian noise areusually the two corruption methods. Variational autoencoder (VAE)
Different from original AEmodel, VAE [75], [76] is a probabilistic generative model.The latent representation h of inputs is not directly learned byencoder, but being encoded learning by encoder to generate adesired latent probabilistic distribution at condition of proba-bilistic constraint. Generally, this constraint is standard normaldistribution, i.e., N(0,1). In phase of decoding, sampling fromthe latent distribution representation h , the decoder generatesthe output. Therefore, the VAE has two loss function: one forencoder to evaluate the similarity between generated distri-bution by encoder and standard distribution, the other is formeasuring how close between the original input and the outputdata. The idea of generator of GANs is also from the VAE.please refer to related literature [33] for other AEs, such as,contractive autoencoder [77], convolutional autoencoder [78].
3) Convolutional Neural Networks (CNNs) .Inspired by animal’s visional neural information process-ing system, CNNs are extensively applied to many researchdomains [38], including CV, NLP, speech recognition. Spe-cialized convolution layer and pooling layer, CNNs canquickly extract latent features of data by shared convo-lutional kernel and downsampling with pooling operation,which characterizes with the positive properties of partiality,identity, and invariance. Up to now, many famous CNNsarchitectures have emerged, (including one-dimension, two-dimension, and multi-dimension, the relative diagrams showedin Fig. 9), such as LeNet-5 [79], AlexNet [80], VGGNet[81], GoogleNetv1-v4 [82]–[85], ResNet [86], MobileNetv1-v3 [87]–[89], ShuffleNetv1-v2 [90], [91], and the latest Ghost-Net [93]. The developments of classic CNNs models are shownin Fig. 8. The increasing depth of model is a main fashion at
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 8
LeNet-5
AlexNet
ZFNet GoogleNet(Inception v1)VGGNetsGAN
ResNet
SqueezeNetInception v2,v3 Inception v4SENetShuffleNet v1DenseNetMobileNet v1XceptionResNeXt ShuffleNet v2MobileNet v2 MobileNet v3 GhostNet1998 2012 2013 2014 2015 2016 2017 2018 2019 2020 (cid:266)(cid:266)(cid:266)(cid:266)
Fig. 8. The classic CNN models [38]. the starting time of DL, e.g, from the starting with a 5-layer(two convolution-pooling layers and three fully connectionlayers) of LeNet in 1998 to hundreds of layers of ResNetin 2015 [86]. In recent years, the lightweight models design,i.e., small volume of parameters, is increasingly popular, e.g.,ShuffleNet, MobileNet, EfficientNet [92].LeNet-5 has been successfully applied to handwritten digitsrecognition [79], which is equipped with two convolution-pooling layers (convolution kernel: ∗ , and ∗ ) and threefully connection layers, but without activation function. Thisstructure pattern was widely used in most of CNNs mod-els. A 8-layer AlexNet has firstly won the championship inImageNet large-scale visual recognition challenge (ILSVRC)in 2012 [80], which quickly expands the intensive researchinterest of deep learning from the industrials and academics.This competition verified that the deeper the model is, thebetter performance will be. The structure of AlexNet has 5-convolution-pooling layers (convolution kernel: ∗ , ∗ ,and ∗ ), 3-fully connection layers, others containing Reluactivation function, dropout. To increase the depth of model,VGGNet of has been proposed for in ILSVRC 2014 [81],won the second place, including VGG-11, VGG-13, VGG-16, VGG-19. Its convolutional kernels are all ∗ , instead of ∗ and ∗ . ! " Fig. 9. The diagram of convolution operation: (a) 1D convolution; (b) 2Dconvolution; and (c) 3D convolution.
The large volume of parameters of deep model, however,leads to low computation efficiency during training process.Combined with multi-parallel filters in the same layer (i.e., Inception module), a 22-layer GoogleNet has won the cham-pion in the competition of ILSVRC 2014 [82], whose numberof parameters is 12 times less than AlexNet, but has higherperformance. GoogleNet firstly verified that the deep modelcan work well by increasing the width of model, not just depth,including GoogleNetv1-v4 [82]–[85]. With the increasing oflayers, the problem of difficult training of model is moreand more obvious, i.e., gradient exploding and vanishing. Toaddress this issue, in [86], the authors proposed a 34-layerof ResNet, which won the ILSVRC 2015 as a championmodel. The excellent design of ResNet is the skip connectionoperation of input to directly output, not through the hiddenlayer. In this way, the model just learns the residual partbetween the ultimate output and original input, which can keepthe gradient existing in a suitable range during all trainingprocess to efficiently train more deeper networks. ResNetmakes extreme deep network possible, such as, ResNet-152.Although ResNet can improve the computation efficiency,a large volume of parameters remains a challenge for op-timally training the model in some practical applications,because of the insufficient computing power and low efficientperformance. In recent years, the lightweight DL modelshave become the main research direction, including the de-sign of lightweight model (such as MobileNets(v1-v3) [87]–[89]), ShuffleNets(v1-v2) [90], [91]), EfficientNet [92]), modelcompression and hardware acceleration technique [41]. Forexample, MobileNets was proposed by Google corporation toembed in portable devices, such as mobile phones.To solve the problem of redundant features extraction ofexisting CNNs, Ghost module was proposed in [93], whichcan be embedded in existing CNNs models to construct a highcomputation efficiency model, i.e., GhostNet, to achieve state-of-the-art performance results in DL tasks.
4) Recurrent Neural Networks (RNNs) .Different from the CNNs, RNNs, inspired by the memoryfunction of animal-based information processing system, areused to solve the problem of data prediction with the temporalmemory series. In other words, the current output results arerelated to previous data sequences. The memory unit, as thebasic module of RNNs, is shown as in Fig. 10. This unitincludes one-layer fully connected neutral network, two input:state s (i.e., the memory of previous unit has m dimensions)and data feature x ( n dimensions), and output as the stateinput of next memory unit. One-layer RNN consists of multi- OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 9 memory units sequently connected, as shown in Fig. 11. Thenumber of memory units is determined by the length ofdata series ( x (0) , x (1) , ..., x ( R − , R is the length of inputsequence). The output of last memory unit is the ultimateresults of RNN learning. All memory units share identicalparameters in the same layer of RNN: weights W of m + n dimensions, bias b of m dimensions. Multiple one-layer RNNstacks to form multi-layer RNN. m x s s Fig. 10. The memory unit in RNNs. m x s s m x s m x (R-1) s R m output Fig. 11. The one layer RNN.
Although the RNNs architecture can achieve the functionof memory, the gradient vanishing issue is obvious with theincreasing length of time series during the training process.To address this problem, long short-term memory (LSTM)architecture is proposed in [94]. Compared to original memoryunit, a LSTM module has two states: long-term memory unit( C ) and short-term memory unit ( h ), both are m dimensions. C can selectively memorize valuable information of long tem-poral series, which efficiently transmit the early information tocurrent unit. LSTM consists of four gate units: forget , memory , information and output , respectively. Each gate unit includes afully connection neural network layer with m neutrons, and theoutput of each gate is short time memory h and data features.The activation functions are sigmoid , except for information gate is tanh function, since the output of sigmoid ranges from0 to 1, contributing to the functions of forget , memory , and output dramatically. The structure of LSTM is shown in Fig.12 and the main relationship is shown as the follow,Firstly, the f orget gate unit determines which kind ofinformation should be discarded from the input f t = sigmoid ( W f [ h t − x t ]) + b f . (17)The following is the memory and inf ormation gate unitsto determine the input of new information, i.e, input gate unit x +x x (cid:169) (cid:169) (cid:55)(cid:68)(cid:81)(cid:75) (cid:169) Tanh h t-1 x t C t-1 C t h t h t f t i t C t x (cid:55)(cid:68)(cid:81)(cid:75) CCCC tt ~ o t Fig. 12. The LSTM module. i t = sigmoid ( W i [ h t − x t ]) + b i . (18) e C t = T anh ( W c [ h t − x t ]) + b c . (19)The new long time memory ( C ) is acquired by C t = C t − ∗ f t + i t ∗ e C t . (20)The output gate unit is e o t = sigmoid ( W o [ h t − x t ]) + b o . (21)Lastly, the short time memory ( h ) is given by h t = o t ∗ T anh ( e C t ) . (22) x + x (cid:169) (cid:169) (cid:55)(cid:68)(cid:81)(cid:75) x t h t-1 h t h t r t z t h t x (cid:55)(cid:68)(cid:81)(cid:75) hhhh tt ~1-x Fig. 13. The GRU module.
Accordingly, LSTM architecture can solve the gradientvanishing issue, thanking to the long time memory unit ( c )and f orget gate unit. f orget gate discards much non-valuableredundant information and c can preserve valuable informationwith large numerical value, therefore, the gradient will notbecome smaller after layer-by-layer gradient decent training,and avoid emerging gradient vanishing to some extent. As a OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 10 variant of LSTM, gated recurrent unit (GRU), which combines f orget gate with input gate (i.e., memory and inf ormation gates aforementioned) as a single update gate, is more simplethan LSTM [95]. The GRU module is shown in Fig. 13. Therelationships of variables are shown as following, r t = sigmoid ( W r [ h t − x t ]) + b r , (23) z t = sigmoid ( W z [ h t − x t ]) + b z , (24) e h t = T anh ( W h [ h t − ∗ r t x t ]) + b h , (25) h t = (1 − z t ) ∗ h t − + z t ∗ e h t . (26)Although the performance of GRU is similar to LSTM, thestructure of GRU is simpler than that of LSTM. The amountof parameters of GRU is only one third of LSTM. Therefore,GRU converges fast and does not cause overfitting.
5) Generative Adversarial Networks (GANs) .Similar to VAE, GANs are also unsupervisedgenerative models, which consist of
Generator G and
Discriminator D [96]–[98], as shown in Fig. 14. The inputof G is usually noise with standard normal distribution, i.e., N (0 , , to generate a new sample (e.g., image) as output.The D is an two-class classifier, to discriminate whetherthe generated sample is true or not. So the inputs are newgenerated sample and true sample, and the output is theprobability of classification. The loss function of D has twoparts: loss , determined by true sample and true labels, and loss is for generated sample and its label. The D makescorrect discrimination between generated and true samplesby minimizing the ( loss + loss ) . The G has just one lossfunction loss determined by generated sample and true label,to try to trick the D . The loss function of GAN is as follow min G max D V ( G, D ) = E x ∼ p data ( x ) [log D ( x )]+ E z ∼ p z ( z ) [log(1 − D ( G ( z )))] , (27)where p data ( x ) , z , p z ( z ) , and G ( z ) represent true distri-bution of sample, noise signal, distribution of noise signal,and generated new sample, respectively. The distribution of G ( z ) is p G ( x ) . D ( x ) and − D ( G ( z )) denote the loss ofdiscriminator and generator respectively.GAN firstly trains discriminator to maximize the expecta-tion of discrimination, which tries to correctly discriminatethe true and generated samples. Then, fix the parameters ofgenerator to minimize the divergence (i.e., Jensen-Shannon(JS) divergence [96]) between the true and generated samples.In other words, the purpose of this phase is making thedistribution of the generated sample close to distribution oftrue sample as close as possible. So the discriminator is usedto measure the gap between the generated and true distribution,instead of directly computing the generated distribution ofgenerator p ( G ( x )) . The training process will not stop untilthe discriminative probability of true and generated sample isequal, i.e., . . Random noise(z) Generator(G) xTrue sample (x)
Discriminator (D)G(z) Loss True/False?
Fig. 14. The diagram of GAN.
Although supervised learning representation with CNNshas developed many achievements in CV domain, the la-beled datasets remains a great challenge. GANs have beendemonstrated huge potentials in unsupervised learning, whichbridges the gap between supervised learning and unsupervisedlearning with deep CNNs architecture. Deep convolutionalGANs (DCGANs) were proposed in [99]. However, GANssuffer from training instability, and it is difficult to adjust thediscriminator to an optimal state.To address this issue, the authors proposed WassersteinGAN (WGAN) model in [100], [101] to make the process oftraining easier by using a different formulation of the trainingobjective that does not suffer from the gradient vanishingproblem. WGAN replaces JS divergence in original GANmodel with Wasserstein distance as objective loss function,which transforms the binary classification into regressionmodel to fit Wasserstein distance. The discriminator of WGANmust satisfy the space of 1-Lipschitz functions, which enforcesthrough weight clipping. The objective of WGAN is as follow[101] min G max D ∈ Ω W ( p r , p g ) = E x ∼ p r ( x ) [ D ( x )] − E z ∼ p g ( z ) [ D ( G ( z ))] , (28)where Ω is the set of 1-Lipschitz functions, p g ( z ) is thedistribution of generator, and p r ( x ) is the true distribution ofsample.Moreover, there are also some other GANs like conditionGAN [102], cycle GAN [103], conditional cycle GAN [104],InfoGAN [105].
6) Reinforcement Learning (RL) .Reinforcement learning (RL) system is also an unsupervisedlearning framework concerning on iteratively making optimaltactical actions act on environment to obtain maximum totalamount of rewards [16]. RL is a Markov decision process(MDP) with the interactions between the artificial agent andcomplex and uncertain environment regarding the sets of statesand actions. The exploration-exploitation trade-off is a typicaltraining processing of RL. The former is to explore the wholespace to aggregate more information while the latter is toexploit the information with more value at the conditions ofcurrent information. As the usual RL algorithm, Q-learning(also action value function) aims to obtain a Q function tomodel the action-reward relationship. Bellman equation is usedto calculate the reward in Q learning. The neural network isoften used to model the Q function in deep Q network (DQN).
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 11
Radar signal automatic modulations recognition(Intra-pulse modulation, PRI modulation)Radar emitters recognitionRadar waveform classification Jamming or interference recognition R a d a r r a d i a t i o n s o u rce s c l a ss i f i c a t i o n a nd rec og n i t i o n (cid:1062) RR S CR (cid:1063) Fig. 15. The contents of RRSCR.
III. R
ADAR R ADIATION S OURCES C LASSIFICATION AND R ECOGNITION
Electronic warfare (EW) is one of the crucial aspects ofmodern warfare [106]. EW receivers are passive systems thatreceive emission from various platforms that operate in therelative vicinity. The received signals are typically analysed[163] to obtain valuable information about characteristicsand intentions of various elements that are presented in thebattlefield. A significant example in modern military warfare isrepresented by radar radiation sources classification and recog-nition (RRSCR) [109], [151], which is one of the tasks thatare associated to electronic support measures and electronicsignal intelligence systems (ESM/ELINT) [107], [108]. Theformer (ESM) focuses on classifying different radar types,such as military or civil radar, surveillance or fire controlradar, whereas the latter further concerns the identification ofindividual radar emitter parameters of the same classification,also called specific emitter identification (SEI) [110], [147],[150], [152]. Such operations are based on radio frequencydistinct native attribute (RF-DNA) fingerprint features anal-ysis methods [111], such as pulse repetition interval (PRI)modulations analysis, intra-pulse analysis. For example, kernelcanonical correlation analysis [146] and nonlinear dynamicalcharacteristics analysis [147] have been used to recognizeradar emitters. In addition, analysis of out-of-band radiationand fractals theory were reported in [148], [149]. Theseradar radiation sources (RRSs) include signal carrier frequency(SCF), linear frequency modulation (LFM), non-LFM, sinu-soidal frequency modulation (SFM), even quadratic frequencymodulation (EQFM), binary frequency-shift keying (2FSK),4FSK, dual linear frequency modulation (DLFM), mono-pulse (MP), multiple linear frequency modulation (MLFM),binary phase-shift keying (BPSK), Frank, LFM-BPSK and2FSKBPSK [153], [215]. In this section, RRSCR includethe classification and recognition of radar signal automaticmodulations (such as intra-pulse modulations, PRI modula-tions), radar emitter types, radar waveforms, and jamming orinterference, as shown in Fig. 15. Examples of time-frequencysamples of RRSs are shown in Fig. 16.RRSCR mainly concerns the following four aspects:i) denoising and deinterleaving (or separation) of collectedpulse streams;ii) improving accuracy of recognition in low SNR scenarios,in conditions of missing and spurious data and in real-time;iii) boosting robustness and generalization of algorithms;iv) identification of unknown radiation sources.The methods of RRSCR mainly have three classes: (cid:11)(cid:68)(cid:12) (cid:11)(cid:69)(cid:12) (cid:11)(cid:70)(cid:12) (cid:11)(cid:71)(cid:12) (cid:11)(cid:72)(cid:12) (cid:11)(cid:73)(cid:12)
Fig. 16. The time-frequency images of RRSs.(a)SCF,(b)LFM,(c)non-LFM,(d)BPSK,(e)2FSK,(f)4FSK. i) knowledge based;ii) statistical modeling based;iii) ML based.The knowledge-based methods depend on the prior radarknowledge summarized from the collected raw data by radarexperts to achieve RESCR-related works. A novel knowledge-related radar emitter database was built by relational mod-eling in [155]. In [108], the authors proposed radar signalknowledge representation determined by rules with semanticnetworks. The authors also analyzed signal parameters, featureextraction using linear Karhunen-Loeve transformation andapplied knowledge-based techniques to recognize the inter-cepted radar signals [154]. Concerning traditional statisticalmodeling methods, an autocorrelation spectrum analysis wasapplied to [156] for modulation recognition of multi-inputand multi-output (MIMO) radar signals. In [157], a jointsparse and low-rank recovery approach was proposed for radiofrequency identification (RFI), i.e., radar signal separation. Inaddition, a feature vector analysis based on a fuzzy ARTMAPclassifier for SEI was developed in [159], a wavelet-basedsparse signal representation technique was defined for signalseparation of helicopter radar returns in [160], and an entropy-based theoretical approach for radar signal classification wasdeveloped in [161].The increasingly growing complexity of electromagneticenvironment demonstrates severe challenges for RRSCR, suchas the increasingly violent electronic confrontation and theemergence of new types of radar signals generally degradethe recognition performance of statistic modeling techniques,especially at low signal noise ratio (SNR) scenario. Althoughthese aforementioned technologies can improve performances,they are not sufficient to face these challenges. Knowledge-based methods spend considerable time to extract signalfeatures. Conventional statistical modeling methods dependon statistical features of the collected data. However, thisoperation pattern do not have competitive performance.In recent years, because of the high-efficiency of ML algo-rithms and the rapid development of novel RSP technology,ML-based methods have been successfully applied to RRSCRto face some critical challenges. To better understand theseresearch developments and grasp future research directionsin this domain, we provide a comprehensive survey on ML-
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 12
TABLE I: The traditional ML algorithms in RRSCR
Features Models Accuracy
PWDs [170], [175], [176], [179];entropytheory [161]; spectrum features [112], [171],[177]; wavelet packets [172]; dynamicparameters searching [173]; rough sets [174];energy envelope [181]; time-frequency analysis[185]–[188]; autocorrelation images [113],[156], [189], [190]; CWTFD [114]–[117];PCA [191], ambiguity function images [125] SVMs [170], [192]–[194], [196]; ANNs [161],[170], [171], [182], [197]–[202]; DT [118],[119], [203]; RF [170]; Adaboost [120];clustering [121]–[124], [161]; K-NN[126]–[129]; weighted-Xgboost [180]; HMMs[204] 84% (-5 dB) [161]; 97.3% (-6 dB) [117]
TABLE II: The DL algorithms in RRSCR
Features Models Accuracy
IQ 1D time sequences [138], [210], [218],[569]; STFT [133]–[135], [137], [212], [229],[230]; CWTFD [130], [215], [217]–[219],[227]; amplitude-phase shift [211]; CTFD[131], [221], [222]; bivariate image with FST[132]; bispectrum [237]; autocorrelationfeatures [213]–[215]; ambiguity functionimages [140], [141]; fusion features [139],[220] CNNs [82], [210], [211], [217]–[222],[228]–[231], [233], [237], [569]; RNNs[142]–[144], [216]; DBNs [135], [136], [235],[236]; AEs [222]; SENet [212], [213];ACSENet [214], [215]; CDAE + CNN[222]–[224]; CNN + DQN [131]; CNN +LSTM [145], [226]; CNN + TPOT [225];CNN + SVM [227] 94.5% (-2 dB) [218]; more than 95% (-9 dB)[222]; 93.7% (-2 dB) [217]; over 96.1% (-6dB) [221]; 96% (-2 dB) [137]; more than 90%(-6 dB) [145]; more than 94% (-6 dB) [131];95.4% (-7 dB) [223]; 94.42% (-4 dB) [225];97.58% (-6 dB) [228].
Preprocessing Deep model design Model trainingRadiation sources EvaluationFeatures selection Classifier design Classifier training
Traditional machine learningDeep learning
RF system
Fig. 17. The pipeline of RRSCR. related RRSCR in this section. This is roughly divided into twoparts: one concerning traditional ML algorithms and the otheris DL-based methods. A concise summary of some examplesof the existing algorithms is shown in Table I and Table IIfor traditional ML and DL-based algorithms, respectively. Ageneric pipeline of ML-based methods is also shown in Fig.17, to represent a visual framework of ML algorithms.
A. Preprocessing
Data preprocessing is the first step, which processes col-lected raw data (i.e., sequence data) to prepare for the fol-lowing classification or recognition tasks, including denoising,deinterleaving [164], data missing processing [170], [182],unbalanced dataset [180], [183], noise and outliers, featuresencoding and transformation [170], [183], [184], and datascaling. we will introduce the denoising, deinterleaving, andfeatures transformation.Because of complex electromagnetic environment, amountof interleaving radio signals are hard to classify and recognizedirectly in short time, so deinterleaving is the first step. Multi-parameter cluster deinterleaving methods are usually adoptedfor deinterleaving the pulse streams (such as pulse repetitioninterval (PRI) deinterleaving methods [165], time of arrival(TOA) deinterleaving methods [166]). Some novel methodshave emerged based on ML algorithms in recent years. Param-eter clustering technology was proposed to deinterleave the receptive radar pulses based on Hopfield-Kamgar [162] andFuzzy ART neural network [163]. To solve the deinterleavingproblems of pulse streams, a group of forward/backwardprediction RNNs was established in [164] to understand thecurrent context of the pulses and predict features of upcomingpulses. The cluster and SVM classifier were employed tointerleave mixed signals with similar pulse parameters in[167]. In [168], MLP structure was used to deinterleave theradar pulse train. As for denoising aspects, RNNs was usedfor denoising the pulse train in [164]. AEs are also used toaddress pulse denoising problem by extracting features fromTOA sequences [169].As for features transformation, the one-dimension and two-dimension features are usually the inputs of DNN models. Theformer are encoded IQ time sequences [138], [210], [218],[569], and the latter usually are time-frequency distribution(TFD) images, which are produced by short time fourier trans-formation (STFT) [212], Choi-Williams time-frequency distri-bution (CWTFD) [217], [218], and Cohen’s time-frequencydistribution (CTFD) image [221], [222]. In addition, there aresome other two-dimension feature images, such as amplitude-phase shift image [211], the spectrogram of the time domainwaveform based on STFT [230], bispectrum of signals [237],ambiguity function images [140], [141], and autocorrelationfunction (ACF) features [213]–[215].
B. Traditional Machine Learning in RRSCR
Traditional ML algorithms based in RRSCR usually in-cludes features selection, classifier design, classifier trainingand evaluation. Two-phase method of feature extraction andclassification based on common machine learning algorithm,is a typical pattern in RRSCR reported in many literatures.There are many classifier models applied to RRSCR, suchas supervised learning methods: ANN [56], SVMs [44], [45],decision DT [48], RF [52], as shown in Table.I.
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 13
The accuracy rate of the traditional two-step methods ismainly determined by the feature extraction algorithm. Artifi-cial feature extraction regarding specific types of radar signals,however, depend mostly on the experience of the experts.Compared to two-step method, DL-based methods can developfeature extraction automatically and potentially learn the latentfeatures of data, so it has higher accuracy. Challenges ongeneralization, big dataset, and optimal training algorithm,however, are main problems for DL-based methods.Feature extraction is used to extract signal features fromthe preprocessed data for classification model training andrecognition [184]. These features include pulse descriptionwords (PDWs) of radar signal [170], [175], [176], [179], infor-mation entropy theory [161], high order spectra [171], waveletpackets [172], dynamic parameters searching [173], rough sets[174], acousto-optic spectrums [177], energy envelope [181],time-frequency analysis [185]–[188], autocorrelative functions[189], [190], principal component analysis (PCA) method[191]. Parameters of signal, however, is time-variable, whichcan lead to uncertainty of signal. Vector neural network wasreported in [178] to deal with the uncertainty of parameters.
1) SVMs classifiers .With the typical advantage of efficiently using kernel func-tion to deal with non-linear binary classification, SVMs aremainstream of ML methods applied to RRSCR [170], [192]–[194], which maximizes the distance or margin between thesupport vectors of classes to search for an optimal hyperplanein feature space of samples.In [170], SVM was used in radar signal classification andsource identification based on the PDWs of radar, includingcontinuous, discrete and grouped radar data signal train pulsesources. To simplify SVM structure and improve recognitionaccuracy, SVM with binary tree architecture was proposed in[192], a roughly pre-classification method was used beforeSVM with resemblance coefficient classifier. Transient energytrajectory-based SVM method was proposed in [193] forspecific emitter identification with robustness to Gaussian-noise, which used PCA to deduce dimensions of featuresspace. To address the non-linear classification, there are lotsof researches on kernel-SVM in RRSCR with different ker-nel functions. However, optimal kernel function is basicallyrelative to excellent performances in stability and accuracy.In [194], the authors developed the comprehensive estimationmethod for choosing optimal kernel functions of SVM forradar signal classifier, which used separability, stability andparameter numbers as evaluation indexes.To identify the radar emitter sources with high accuracyrate at low SNR scenario, a SVM classifier based on thescale-invariant feature transform (SIFT) in position and scalefeatures was employed in [196]. The SIFT scale and positionfeatures of the time-frequency image were extracted based onthe Gaussian difference pyramid. The extracted noise featurepoints were suppressed based on the scale features. Finally,SVM was used for the automatic identification of radiationsources based on the SIFT position features.However, SVM classifier does not good at learning newknowledge in real-time. Hull vector and Parzen window den-sity estimation [195] were reported for online learning of radar emitter recognition.
2) Artificial Neural Networks (ANNs) classifiers .This part will review ANNs-based methods popularly ap-plied to RRSCR, only considering superficial layer NNs,which have not more than 3 hidden layers, including vectorneural network [197], [198], SPDS-neural network [199],radial basis function neural network (RBF-NN) [200], andfusion neural network [201]. The DNNs-based related workswill be surveyed in later section.To guarantee the accuracy rate of approximately 100% inexacting one-dimensional parameter, a modified back propa-gation (BP) neural network was proposed in [199] for radaremitter recognition with uncomplicated data and enough traintime. RBF-NN was developed in [200] to classify radar emittersignals. The decision rules of RBF-NN, to determine signaltypes, are extracted from rough sets and the cluster centerof RBF-NN by rough K-means cluster method. A what-and-where neural network architecture was developed for recog-nizing and tracking multiple radar emitters in [201].The multi-layer perceptron (MLP) achieved more than 99%recognition rate at SNR ≥ OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 14 a classifier to classify instantaneous transient signals based onRF fingerprint extraction [181]. As for unknown radar signalsrecognition, class probability output network (CPON) wasproposed in [208] for classification of trained and untrainedradar emitters signal types.When classifing intercepted radar signals, there exists a datadeviation in practical application. Weighted-xgboost algorithmwas applied to [180] to address this problem. Compared toexisting methods, this novel method achieved 98.3% accuracyrate, while the SVM, RVM [206], the gradient boost methods,and DBN obtained 89.1%, 79.6%, 91.9%, 95.4%, respectively.In [209], AdaBoost algorithm was developed as a classifierbased on fast correlation-based filter solution (FCBF) featureselection, to complete the recognition of different types ofradar signals with 1D harmonic amplitude datasets. Thesedatasets were decomposed by frequency-domain analysis fromradar time-domain signals. The simulation results showed thatthis method was more effective than the SVM algorithm inaccuracy rate and stability.
C. Deep learning in RRSCR
Compared to statistics-based analysis methods, traditionalML-based have developed many achievements in RRSCR in-troduced in section B, which can improve the classification andrecognition performance dramatically. However, the weaknessof standard 2-phase-method is hard to further extract latentfeatures by domain experts to facilitate classification modeltraining, because of the limitation of expert knowledge andlots of time costs in general.Nowadays with the advantages of deeply automatic featureextraction, radar experts exploit apply DL in RRSCR toimprove the classification performance based on DNN models.In general, the 1D and 2D features are as the inputs ofDNN models aforementioned. Since the CNNs have excellentperformance and have been applied widely to image classi-fication and recognition. In this section, we mainly make acomprehensive survey on radar signals classification based onCNNs architecture. In addition, RNNs [216], DBNs [235],[236], and AEs [222] are also briefly investigated.A novel unidimensional convolutional neural network (U-CNN) was proposed in [210] to classify radar emitters,which is based on encoded high dimension sequences asextracted features. The U-CNN has three independent convo-lution parts followed by a fully-connected part. Three encodedsequences: RF i , P RI i , P W i act as inputs of the correspond-ing convolution parts. Experiments on a large radar emitterclassification (REC) dataset, including 67 types of radars and227,843 samples, demonstrated that U-CNN can achieve thehighest accuracy rate and competitive computation cost forclassification, compared with other classifier models, such asNN, SVM, DT.A CNN model with five convolution-maxpooling layers, twofully connection layers, and one softmax output layer, wasproposed in [211] to classify radar bands from mixed radarsignals. Experiments results showed that amplitude-phase shiftproperty as inputs of CNN achieved 99.6% of accuracy rate,compared to that of 98.6%. when spectrograms as inputs.Sequeeze-and-excitation network (SENet) was proposed in [212] to identify five kinds of radar signals, each of which has4,000 training samples. This novel model achieved accuracyrate of 99% with time, frequency, and TFD images as theinputs. Combining with autocorrelation functions, SENet wasalso used in [213] to recognize PRI modulations. Moreover, in[214], asymmetric convolutional squeeze-and-excitation net-work (ACSENet) and autocorrelation features were proposedfor PRI modulations. Also, in [215], multi-branch ACSENetand multi-dimension features based on SVM fusion strategywere developed for multiple radar signal modulation recogni-tion. Similarly, a CNN model was employed in [229] basedon multiple zero-means scaling denoised TFD images of radaremitter intra-pulse modulated signals.A cognitive CNN model was proposed in [217] to recognize8 kinds of radar waveforms based on CWTFD-based TFDimages. More than probability of successful recognition (PSR)of 90% was achieved when the SNR was -2 dB. To improvethe accuracy rate, an automatic radar waveform recognitionsystem was exploited in [218] to detect, track and locate theLPI radars. This novel system achieved overall PSR of 94.5%at an SNR of -2 dB by a hybrid classifier. The model includestwo relatively independent subsidiary networks, mainly CNNand Elman neural network (ENN) as auxiliary.In [219], the authors proposed a deep CNN based automaticdetection algorithm for recognizing radar emitter signals,which leveraged on the structure estimation power of deepCNN and the CWTFD-based TFD images as inputs of model.This architecture had competitive performance compared withBP and SVM models. Combining CNN model with the newkernel function, CTFD as the inputs of model for identifying12 kinds of modulation signals to achieve more than PSR of96.1% at the SNR of -6 dB [221].To make full use of the features of inputs, a feature fusionstrategy based on CNN architecture was proposed in [220]to classify intra-pulse modulation of radar signals with fusedfrequency and phase features. Two independent CNNs learnedfrequency and phase related inputs respectively, and thenfollowed by feature fusion layer to fuse the individual outputsas ultimate output. Similarly, two different neural networkswere developed in [230] with spectrogram of the time domainwaveform by STFT for radar emitter recognition.In order to accelerate feature learning of CNN, a PCAbased CNN architecture was proposed in [231] to reducedimensionality of TFD images. After feature extraction withCNN, random vector functional link (RVFL) was employed in[233] to promote feature learning ability, and picked out themaximum of RVFL as identification results of signals.In general, TFD images remove noise by preprocessingprocess before them are as inputs of CNN, such as binariza-tion and wiener filtering [220], [221], [229]. Although thispreprocessing pattern can reduce the impact of noise, it maycause a loss of information details contained in images to someextent. To address this problem, an end-to-end DL methodwas developed in [222] to recognize 12 classes of intra-pulse modulation signals based on convolutional denoisingautoencoder (CDAE) and deep CNN with CTFD-based TFDimages. CDAE was used to denoise and repair TFIs, andInception [82] based deep CNN was used for identification. OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 15
The simulations showed that the proposed approach had goodnoise immunity and generalization and achieved PSR of morethan 95% at SNR of -9 dB for twelve kinds of modulationsignals classification. An end-to-end RNN architecture wasproposed in [216] for classification, denoising and deinterleav-ing of pulse streams. This structure used RNNs to extract longterm patterns from previous collected streams by supervisedlearning and understand the current contexts of pulses topredict features of upcoming pulses.Pulse repetition interval (PRI) is a vital feature parameterof radar emitter signals. It is possible to recognize radaremitter only based on PRI of signals. Due to the high ratio oflost and spurious pulses in modern complex electromagneticenvironments, however, PRI modulations are more difficult toseparate and recognize. To address this issue, A CNN modelwas proposed in [237] to recognize the PRIs modulations ofradar signals. Simulation results showed that the recognitionaccuracy is 96.1% with 50% lost pulses and 20% spuriouspulses in simulation scenario.A more efficient threat library was generated in [234] forradar signal classification based on DBN model, consistedof independent RBMs of frequency, pulse repetition interval,pulse width respectively, and a RBM fused the pervious resultsagain. The experiments results showed more than 6% perfor-mance improvement over the existing system. To accuratelyaddress the complex electromagnetic environment and varioussignal styles, a robust novel method based on the energyaccumulation of STFT and reinforced DBN was developedin [235] to recognize radar emitter intra-pulse signals at alow SNR. Deep network based hierarchical extreme learningmachine (H-ELM) was explored in [236] for radar emitter sig-nal representation and classification with high order spectrum.After extracting the bispectrum of radar signals, the SAE inH-ELM was employed for feature learning and classification.Although DL has high accuracy and generalization forRRSCR, its black-box property makes it difficult to apply inpractical applications, such as military and medical applica-tions. To alleviate this issue, a novel method was presented in[232] based on tree-based pipeline optimization tool (TPOT)and local interpretable model-agnostic explanations (LIME).The experimental results showed that the proposed methodcan not only efficiently optimize the ML pipeline for differentdatasets, but determine the types of indistinguishable radarsignals in the dataset according to the interpretability.In summary, this subsection has done a comprehensivesurvey on the RRSCR based on ML algorithms, includingthe classification and recognition of radar signal modulations,LPI waveform, and radar emitters. The ML algorithms includetraditional ML and DL, such as SVM, DT, adaboost, CNN,RNN, AE, DBN. The features include statistic, 1D, 2D, andfusion features.IV. R
ADAR I MAGES P ROCESSING
Active radar imaging is an important tool for detectionand recognition of targets as well for the analysis of naturaland man-made scenes. Radar images in a broader senseinclude unidimensional high-resolution range profiles (HRRP) [509], [510], [512], [523], two-dimensional SAR and ISARimages [273]–[275], [279], micro-doppler images [551]–[555]and range-doppler images [556]–[558]. Several ML-basedtechniques have been developed for radar image process-ing, particularly for what concerns Synthetic Aperture Radar(SAR) and Inverse Synthetic Aperture Radar (ISAR). Thissection will review the scientific literatures that focus onradar image processing based on ML technology, includingimage preprocessing (e.g., denoising), feature extraction andclassification.
A. SAR Images Processing
Operating conditions of all weather, day-and-night and high-resolution imaging, synthetic aperture radar (SAR) is a popularresearch domain on remote sensing domain in military andcivil applications. SAR is an active remote sensor, i.e., itcarries its own illumination and does not depend on sunlightlike optical imaging. With the rapid development of militaryand science technology, various types of SAR sensors havebeen appeared, which can be roughly divided into three maincategories based on the carrier platform: satellite-borne SAR,airborne SAR, and ground-based SAR. Different SAR sensorscan have different configured properties, even though withinthe same category, such as carrier frequency/wavelength,imaging mode (e.g., stripmap SAR, spotlight SAR, scanSAR,inverse SAR, bistatic SAR and interferometric SAR (InSAR)),polarization (e.g., horizontal (H) or vertical (V) polarization),resolutions in range and azimuth directions, antenna dimen-sions, synthetic aperture, and focusing algorithm (e.g., rangedoppler algorithm, chirp scaling algorithm, and the SPECANalgorithm).Focused SAR image is the 2D high resolution image, i.e.,range and azimuth directions. At range direction, SAR trans-mits LFM waveform with huge product of pulse width andbandwidth, and obtains high resolution of the range directionby adopting pulse compression technology; as for azimuth, along synthetic aperture, formed along the trajectory of relativemotion between detected target and radar platform, to store themagnitude and phase of successive radar echoes to guaranteethe high resolution at azimuth direction. Therefore, one of thevital conditions of forming SAR image is that there shouldexist relative motion between target and radar platform.The multiple configurations of SAR potentially characterizethe distinctiveness of SAR imagery, which vastly contributesto classification and recognition of targets. Compared to op-tical counterparts, SAR images have distinctive characteristicsincluding i) an invariant target size with the various distancebetween the SAR sensor and the target, ii) the imaging sceneinformation is determined by the magnitude and phase of theradar backscatter (i.e., for a single-channel SAR and multi-channel SAR), iii) high sensitivity to the changes of target’spostures and configurations such as the shadowing effect, theinteraction of the target’s backscatter with the environment(e.g., clutter, adjacent targets, etc.), projection of the 3-Dscene (including the target) onto a slant plane (i.e., SAR’sline of sight (LOS)), and the multiplicative noise (known asspeckle) due to the constructive and destructive interferenceof the coherent returns scattered by small reflectors within
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 16 each resolution cell [239], and iv) SAR imagery can easilyobserve the hidden targets with the well penetration of suitablewavelength of electromagnetic wave.SAR imagery processing methodologies includes denoising,classification and recognition, detection and segmentation. Inrecent years, with the rapid development of ML in imageprocessing, ML-based, especially DL, has applied to SARimage processing widely and successfully (such as [251],[252], [273], [297], [343], [493]). In this section, we make acomprehensive survey for SAR image processing techniquesbased on DL algorithms, such as CNN, DBN, SAE.
1) Datasets and Augmentation . Datasets
Dataset is one of the important factors for thesuccess of DL, including training datasets, validation datasets,and testing datasets, respectively. The collection of data andthe building of formatted datasets are challengeable tasks,generally requiring huge human and economic costs. Sinceespecially military backgrounds, the public big SAR datasetsare not easily collected, compared with general CV datasets,such as ImageNet, COCO, CFAR-10, which depends on bigdata easily collected from the Internet. Luckily, with thecooperation and endeavor of radar community, there are stillsome public SAR datasets in military and civil applicationfor target classification and recognition, detection, and seg-mentation. These targets include military vehicles, farmland,urban streets, and ships. Such as moving and stationarytarget acquisition and recognition (MSTAR) [238], [239],[245] , TerraSAR-X high resolution imagery [246], [298],San Francisco [240], Flevoland [240]–[244]. Ship datasets in-cludes SSDD [247], SAR-Ship-Dataset [248], AIR-SARShip-1.0 [249], HRSID [250].MSTAR is a typical widely applied as a baseline SARimagery dataset, including 10 classes of ground targets. Thedataset consists of X-band SAR images with 0.3 m * 0.3 mresolution of multiple targets, which includes BMP2 (infantrycombat vehicle), BTR70 (armored personnel carrier), T72(main tank), etc. All images are size of 128 * 128. The samplesare as shown in Fig. 18.SSDD dataset [247] includes 1,160 images and 2,456 shipstotally, which follows a similar construction procedure asPASCAL VOC [260]. SAR-Ship-Dataset [248] constructedwith 102 Chinese Gaofen-3 images and 108 Sentinel-1 images.It consists of 43,819 ship chips of 256 pixels in both rangeand azimuth directions. These ships mainly have distinct scalesand backgrounds. It can be used to develop object detectorsfor multi-scale and small object detection.AIR-SARShip-1.0 [249] firstly released 31 images, scale of3,000 * 3,000 pixels, the resolution of SAR images is 1 mand 3 m, imaging pattern including spotlight mode, stripemapmode, and single polarization mode. The landscapes includingport, island, the sea surface with different sea conditions.The targets have almost thousands of ships with ten classes,including transport ship, oil ship, fisher boat, and so on.High resolution SAR images dataset (HRSID) [250] is usedfor ship detection, semantic segmentation, and instance seg-mentation tasks in high-resolution SAR images. This datasetcontains a total of 5,604 high-resolution SAR images and16,951 ship instances. ISSID draws on the construction pro- ( a ) ( b ) ( c ) ( d ) ( e ) ( f ) ( g ) ( h ) ( i ) ( j ) Fig. 18. The MSTAR data samples, optical images (top) and theircorresponding SAR images (bottom). (a)2S1, (b)BMP2, (c)BRDM2,(d)BTR60, (e)BTR70, (f)D7, (g)T62, (h)T72, (i)ZIL131, (j)ZSU234. cess of the Microsoft common objects in context (COCO)dataset, including SAR images with different resolutions,polarizations, sea conditions, sea areas, and coastal ports.This is a benchmark dataset for researchers to evaluate theirapproaches. The resolution of ISSID is 0.5 m, 1 m, and 3 m.
Data augmentation
Although there exist some public avail-able datasets, the number of labeled samples is relatively small,which do not always satisfy the requirements of DL algorithm.Therefore, the SAR targets recognition and classification canbe regarded as small samples recognition problem. To addressthe deficiency samples of datasets, many researchers haveproposed novel methods to augment the dataset, such as GANs[251], [252], or design novel efficient model to learn withlimited labeled data, such as TL based methods [252], [253],[256].Wasserstein GAN, with a gradient penalty (WGAN-GP),was proposed to generate new samples based on existingMSTAR data in [251], which can improve the recognition ratefrom 79% to 91.6%, from 57.48% to 79.59%, for three-classand ten-class recognition problem, respectively, compared tooriginal MSTAR. In [252], the authors proposed least squaresgenerative adversarial networks (LSGANs) combined withTL for data augmentation. Different from [251], [252], someimage processing methods were utilized in [253] i.e., manual-extracting sub-images, adding noise, filtering, and flipping, toproduce new samples based on the original data. In [257], theauthors generate noisy samples at different SNRs, multireso-lution representations, and partially occluded images with theoriginal images, to enhance the robustness of CNN at variousextended operating conditions (EOCs). In addition, three typesof data augmentation based on MSTAR were developed in[258], [310], i.e., translation of target, adding random specklenoise to the samples, and posture synthesis. Image reconstruc-tion with sparse representation was proposed in [254], [307],[327] for data augmentation based on attributed scattering
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 17 centers (ASCs).A accuracy-translation map based on domain-specific dataaugmentation method was developed in [311], which canachieve a state-of-the-art classification accuracy rate of 99.6%on MSTAR dataset. In [325], the authors used a flexible meanto generate adequate multi-view SAR data with limited rawdata. An electromagnetic simulation approach was proposedin [331] as an alternative to generate enough bistatic SARimages for network training. In [448], amplitude and phaseinformation of SAR image was also used to generate multi-channel images as the inputs of CNN model to alleviate theover-fitting during the training phase.Except for data augmentation methods, high efficient clas-sification model design is also adopted to alleviate the smallsamples challenge. In [255], the authors proposed a new deepfeature fusion framework to fuse the feature vectors, whichwere extracted from different layers of the model based on Ga-bor features and information of raw SAR images. A TL basedmethod was employed in [256] to transfer knowledge learnedfrom sufficient unlabeled SAR scene images to labeled SARtarget data. A-ConvNet was proposed in [259] to vastly reducethe number of free parameters and the degree of overfittingwith small datasets. The novel architecture replaced the fully-connected layers with sparsely-connected convolutional layers,which obtained average accuracy of 99.1% on classification of10-class targets on MSTAR dataset.
2) SAR Images Denoising .As a coherent imaging modality, SAR images are oftencontaminated by the multiplicative noise known as speckle,which severely degrades the processing and interpretation ofSAR images. It is hard to balance performances betweenspeckle noise reduction and detail preservation. In general,traditional despecking methods [277], [278], (such as multi-look processing [262], filtering [263], [264], [282], [284],blocking matching 3D (BM3D) [265], [280], wavelet-based[266]–[269], [282], separated component-based [270]), trans-form the multiplicative noise into additive noise by logarithmoperation of observed data. These methods, however, canintroduce more or less bias into denoised image. In addition,the local processing of these methods fails to preserve sharplyuseful features, e.g., edges, texture, detailed information, andoften contains artifacts [271]. Another problem is that mostof traditional methods require statistics modeling. To addressthe problems of SAR image despeckling aforementioned,and inspired by the advantages of DL algorithm. DL-basedalgorithm has been applied to this field, especially CNN-basedmodel algorithm.
CNN-based supervised methods
A residual learning strat-egy with residual CNN model was firstly employed in [273]for SAR imagery despeckling, which achieved better perfor-mance on man-synthetic and real SAR data and guaranteeda faster convergence in the presence of limited training data,compared to state-of-the-art techniques. A probability transi-tion CNN (PTCNN) was proposed in [274] to increase noise-robustness and generalization for patch-level SAR image clas-sification with noisy labels. The authors in [275] developed anend-to-end learning architecture (i.e., image despeckling con-volutional neural network (ID-CNN)) to automatically remove speckle from noisy SAR images. In particular, this architecturecontained a component-wise division-residual layer with skip-connection to estimate the denoised image. Similarly, in [292],the authors proposed a SAR dilated residual network (SAR-DRN) to learn a non-linear end-to-end mapping between thenoisy and clean SAR images. DRN could both enlarge thereceptive field while maintaining the filter size and layer depthwith a lightweight structure to conduct image details andreducing the gradient vanishing problem. In addition, com-bined with ensemble learning method, the authors proposed adespecking CNN architecture in [272].To deal with the random noisy SAR imagery, despecklingand classification coupled CNNs (DCC-CNNs) was proposedin [276], to classify ground targets in SAR images with strongand varying speckle. DCC-CNNs contained a despeckling sub-network to firstly mitigate speckle noise, and a classificationsubnetwork for noise robustness learning of target information,which could achieve more than 82% of overall classificationaccuracy rate for ten ground target classes at various specklenoise levels. A novel method to directly train modified U-Net[287] with given speckled images was developed in [286].An extra residual connection in each convolution-block of U-Net, and the operations of replacing the transposed convolutionwith parameter free binary linear interpolation were alsointroduced. A DNN based approach was proposed in [294] forspeckle filtering, which based on DNN’s application in super-resolution reconstruction, iteratively improved the first lowresolution filtering results by recovering lost image structures.To overcome the problem of collecting a large number ofspeckle-free SAR images, a CNNs-based Gaussian denoiserwas developed in [285], which was based on multi-channellogarithm and Gaussian denoising (MuLoG) approaches. TheTL-based pre-trained CNN models, trained by datasets withadditive white Gaussian noise (AWGN), was also directlyemployed to process SAR speckle.Thanks to the excellent ability of exploiting image self-similarity with nonlocal methods, a CNN-powered nonlocaldespeckling method was investigated in [288] to improvenonlocal despeckling performance [289] on man-synthetic andreal SAR data. The trained CNN was used to discover usefulrelationships among target and predictor pixels and wereconverted into the weights of plain nonlocal means filtering. In[293], the authors proposed CNN model combined with guidedfiltering based fusion algorithm for SAR image denoising. Fivedenoised images were firstly obtained via a seven-layer CNNdenoiser acts on an noisy SAR image, then a final denoisedimage is acquired by integrating five denoised images with aguided filtering-based fusion algorithm.However, the DL model remains very sensitive to the inputs.To address the non-invariant denoising capability of DL-basedmethods, a novel automatical two-component DL network withtexture level map (TLM) of images was proposed in [291] toachieve satisfactory denoising results and strong robustness forSAR imagery invariant denoising capability. Texture estima-tion subnetwork produced the TLM of images. Noise removalsubnetwork learned a spatially variable mapping between thenoise and clean images with the help of TLM.
DNN-based unsupervised methods
Except for CNN-based
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 18 supervised learning model methods, the DNN-based unsuper-vised methods are also developed in SAR images denoising.To solve the notorious problem of gradients vanishing andaccelerate training convergence, a AE model was employedin [295] to denoise multisource SAR images, which adoptedresidual learning strategy by skip-connection operation. AE-CNN architecture was developed in [279] for InSAR imagesdenoising in the absence of clean ground truth images. Thismethod can reduce artefact in estimated coherence throughintelligent preprocessing of training data. To solve the trade-offof speckle suppression and information preservation, a CNN-based unsupervised learning solution scheme was proposed in[281], [282]. Taking into account of both spatial and statisticalproperties of noise, this model could suppress the noise whilesimultaneously preserve spatial information details by a novelcost function. In addition, a MLP model was elaborated in[296] for SAR image despeckling by using a time series ofSAR images. The greatest advantage of MLP was that once thedespeckling parameters were determined, they can be used toprocess not only new images in the same area, but also imagesin completely different locations.
3) SAR Automatic Target Recognition (SAR-ATR) .Originated from the military, the goal of ATR is to infer orpredict the classes of detected targets via acquired sensory datawith computer processing technology. Today, ATR technologyis significantly applied to both military and civil domains,such as valuable military target recognition (e.g., missile,airplane, ship), human pose, gait, and action recognition. Dueto the unique characteristics of SAR images aforementioned,it is difficult to easily interpret the SAR imagery with thecommon ATR system. Research on SAR-ATR system hasincreasingly absorbed attention from the researchers aroundthe RSP community. The problems of SAR-ATR researchdomain based on ML algorithms mainly focus on improvingperformances in the following four aspects: i) accuracy withDNNs model, ii) generalization with limited labeled data[343], iii) robustness with speckle denoising, scale-varianceand adversarial samples attack, and iv) real-time or alleviatingcomputation cost at practical strict situations. Furthermore,interpretability of deep models are also studied. SAR-ATR hastwo main categories based on ML: traditional ML based meth-ods, such as SVM [304], [314]–[316], genetic programming[317], boosting [318], Markov random field (MRF) [319]–[322], ELM [338], and DNN based methods, such as CNNs[310], [311], DBNs [312], SAE [313], RNNs [326]. Besides,three classes of SAR-ATR methods have been categorizedin a surveyed paper [34], i.e., feature-based, semi-model-based, and model-based, respectively. This section presentsan understanding survey for SAR-ATR based on model-basedDL algorithms, which is roughly categorized into four classesbased on research aspects aforementioned. These state-of-the-art algorithms including basic CNNs, fusion models, high-way model, multi-view, multi-task learning networks models,RNNs based spatial SAR image sequences learning, AEs, andDBNs. i) Boosting Accuracy with DL ModelGeneral DL models
Similarly, CNNs are also widely ap-plied to SAR-ATR. To understand the relationship between the convolution layers and feature extraction capability, a weightedkernel CNN (WKCNN) was presented in [345]. By modelingthe interdependence between different kernels, this modelintegrated a weighted kernel module (WKM) into the commonCNN architecture to improve the feature extraction capabilityof the convolutional layer. The CNN models were designed forMSTAR data [324] and polarimetric Flevoland SAR dataset(15 classes) [299] classification, which achieved recognitionaccuracy of 99.5% and 92.46% respectively. In [451], the au-thors proposed a dual channel feature mapping CNN (DCFM-CNN) for SAR-ATR, which achieved a average recognitionaccuracy of 99.45% on MSTAR. In order to extract spatialdiscriminative features of SAR images, a DCNN was proposedto extract gray level-gradient co-occurrence matrix and Gaborfeatures in [418], [419] for SAR image classification. Anovel neighborhood preserved DNN (NPDNN) was proposedin [350] to exploit the spatial relation between pixels by ajointly weighting strategy for PolSAR image classification. Anconvolution kernel of the fire module based effective max-fireCNN model, called MF-SarNet, was constructed in [351] foreffective SAR-ATR tasks.Combing DNN and traditional ML algorithm is also in-vestigated. A unsupervised discriminative learning methodbased on AE and SVM models was proposed in [349],called patch-sorted deep neural network (PSDNN), whichfirstly adopted sorted patches based on patch-sorted strategyto optimize CNN model training for extracting the high-levelspatial and structural features of SAR images, and a SVMclassifier as the final classification task. The combination ofCNN and SVM was developed in [357], [358]. A modifiedstacked convolutional denoising auto-encoder (MSCDAE) wasproposed in [359] to extract hierarchical features for complexSAR target recognition, and SVM as final object classificationwith features extracted by MSCDAE model. To enhance thelearning of target features, a novel deep learning algorithmbased on a DCNN trained with an improved cost function,and combined with a SVM was proposed in [360] for SARimage target classification. A TL based pre-trained CNN wasemployed to extract learned features in combination with aclassical SVM for SAR images target classification in [361].Deep kernel learning method was employed in [362] for SARimage target recognition, which optimized layer by layer withthe parameters of SVM and a gradient descent algorithm. Anovel oil spill identification method was proposed in [355]based on CNN, PCA, and RBF-SVM, which could improvethe accuracy of oil spill detection, reduce the false alarm rate,and effectively distinguish an oil spill from a biogenic slick.To take the advantage of manifold learning with modeling corevariables of the target, and separate different data’s manifoldas much as possible, the authors proposed nonlinear manifoldlearning integrated with FCN for PolSAR image classificationin [378].In [354], the authors proposed an ensemble transfer learningframework to incorporate manifold polarimetric decomposi-tions into a DCNN to jointly extract the spatial and po-larimetric information of PolSAR image for classification.In order to effectively classify single-frequency and multi-frequency PolSAR data, the authors proposed a single-hidden
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 19 layer optimized Wishart network (OWN) and extended OWN,respectively in [356], which outperformed DL-based architec-ture involving multiple hidden layers. To exploit the spatialinformation between pixels on PolSAR images and preservethe local structure of data, a new DNN based on sparsefiltering and manifold regularization (DSMR) was proposedfor feature extraction and classification of PolSAR data in[364]. In [380], the authors made full use of existing expertknowledge to construct a novel deep learning architecture fordeep polarimetric feature extraction, and a superpixel map wasused to integrate contextual information. This model consistedof multiple polarimetric algebra operations, polarimetric targetdecomposition methods, and CNN to extract deep polarimetricfeatures.A 20-layers (with 3 dense block and 2 transition layers)DenseNet was built to implement polarimetric SAR imageclassification in [371]. Inception v3 model was adopted in[372] to develop an efficient and accurate method to detectand classify key geophysical phenomena signature among thewhole sentinel-1 wave mode SAR dataset. In [373], the authorsproposed an end-to-end framework for the dense, pixel-wiseclassification of GF-3 dual-pol SAR imagery with convolu-tional highway unit network (U-net) to extract hierarchicalcontextual image features. To concern about the estimationof depression angle and azimuth angle of targets in SAR-ATR tasks, the authors proposed a new CNN architecture withspatial pyramid pooling(SPP) in [385], which could build highhierarchy of features map by dividing the convolved featuremaps from finer to coarser levels to aggregate local featuresof SAR images.To address the redundant parameters and the negligenceof channel-wise information flow, group squeeze excitationsparsely connected CNN was developed in [346]. Groupsqueeze excitation performed dynamic channel-wise featurerecalibration with less parameters, and sparsely connectedCNN demonstrated the concatenation of feature maps fromdifferent layers. This model achieved accuracy rate of 99.79%on MSTAR, outperformed the most common skip connectionmodels, such as ResNet and densely connected CNN. TL-based pre-trained ResNet50 and VGGNet model were em-ployed in [305], [306] for SAR image classification. Pre-trained models can deeply extract multiscale features of datasamples in short training time, and convolutional predictorwere added after the pre-trained model for the target clas-sification. The experiment results showed that the pre-trainedmodel achieved accuracy rate of 98.95% in [305] and higherperformance with the suitable data augmentation technologythan other methods in [306]. TL was developed in [309] toovercome the problem of difficulty in convergence.Instead of directly outputting the class of SAR image withthe DNN, a class center metric based method with CNN modelwas proposed in [333]. This method used CNN to extractfeatures from SAR images to calculate class center of eachclass under the new features representation. Then, the class oftest sample was identified by the minimum distance betweenthe center of class and learned features space of test sample.Similarly, a DNN model was employed in [450] to directlyclassify targets with slow-time and fast-time sampled signals. The decision-making strategy of classification is determinedby the distance between the optimized sets of vectors andclasses. Each of class represented a new sample.As for complex SAR imagery, the complex-value CNN(CV-CNN) architecture was proposed in [300], [301]. Allcomponents of CNNs were extended to the complex domain.CV-CNN achieved accuracy rate of 95% on Flevoland dataset[300], and 96% with enough samples in [301]. Moreover,a deep FCN was also employed in [302] that used real-valued weight kernels to perform pixel-wise classification ofcomplex-valued images. A CV-CAE was proposed in [303] forcomplex PolSAR images classification. In order to sufficientlyextract physical scattering signatures from PolSAR and ex-plore the potentials of different polarization modes on this task,a contrastive-regulated CNN was proposed in the complexdomain, attempting to learn a physically interpretable deeplearning model directly from the original backscattered datain [379]. A novel deep learning framework, deep SAR-Net,was constructed in [377] to take complex-valued SAR imagesinto consideration to learn both spatial texture information andbackscattering patterns of objects on the ground.To exploit the performance of generative models in SAR-ATR based on unsupervised learning, an SAE model withfeature fusion strategy was adopted in [339] for SAR targetrecognition. The local and global features of 23 baselinesand three patch local binary pattern (TPLBP) features wereextracted from the SAR image, which achieved an classifi-cation accuracy rate of 95.43% on MSTAR. A single-layerCNN model combined with features extraction by SAE wasdeveloped in [313], which achieved accuracy rate of 90.1%and 84.7% for 3-class and 10-class targets classification onMSTAR. A novel framework for PolSAR classification basedon multilayer projective dictionary pair learning (MPDPL) andSAE was proposed in [384]. To learn more discriminativefeatures of SAR images, an ensemble learning based discrim-inant DBN (DisDBN) was proposed in [312] to learn high-level discriminant features of SAR images for classification.Some weak classifiers were trained by several subsets ofSAR image patches to generate the projection vectors, whichwere then input into DBN to learn discriminative featuresfor classification. An unsupervised deep generative network-poisson gamma belief network (PGBN) was proposed toextract multi-layer feature from SAR images data for targetsclassification tasks in [352]. An unsupervised PolSAR imageclassification method using deep embedding network-SAEswas built in [353], which used SVD method to obtain low-dimensional manifold features as the inputs of SAEs, and theclustering algorithm determined the final unsupervised classi-fication results. As for In-SAR data, a DBN was used to modeldata in [366] for classification, which could fully explore thecorrelation between intensity and the coherence map in spaceand time domain, and extract its effective features. Inspiredby DL and probability mixture models, a generalized gammadeep belief network (g-DBN) was proposed for SAR imagestatistical modeling and land-cover classification in [383].Firstly, a generalized Gamma-Bernoulli RBM (gB-RBM) wasdeveloped to capture high-order statistical characterizes fromSAR images. Then a g-DBN was constructed to learn high-
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 20 level representation of different SAR land-covers. Finally, adiscriminative network was used to classify. In addition, thedeep RNN based model was adopted in [386] for agriculturalclassification using multitemporal SAR Sentinel-1.
Multi-aspect fused learning methods
Multi-aspect fusedlearning methods are very popular in improving the accuracyof SAR-ATR tasks such as multi-view, multi-task, multi-scale,multi-dimension. To fully extract features of images, a CNNbased fusion framework was proposed in [308], including apreprocessing module initialized with Gabor filters, an im-proved CNN and a feature fusion module. This model couldachieve an average accuracy rate of 99% on MSTAR, evenobtained a high recognition accuracy on limited data and noisydata. The authors developed concurrent and hierarchy targetlearning architecture in [344]. Three CNN models simultane-ously extracted features of SAR images in two scenarios, finalclassification was finished by different combination and fusionapproaches based on extracted features. Based on multi-viewlearning manner, the authors proposed a multi-input DCNN forbistatic SAR-ATR system in [331]. A multi-stream CNN (MS-CNN) was proposed in [343] for SAR-ATR by leveraging SARimages from multiple views. A Fourier feature fusion frame-work derived from kernel approximation based on randomFourier features, to unravel the highly nonlinear relationshipbetween images and classes, which fused information frommulti-view of the same target in different aspects.To capture the spatial and spectral information of a SARtarget simultaneously with kernel sparse representation (KSR)technology, multi-task kernel sparse representation frameworkwas developed in [337] for SAR target classification. SARtarget recognition was formulated as a joint covariate selectionproblem across a group of related tasks. A multi-task weightoptimization scheme was developed to compensate for theheterogeneity of the multi-scale features and enhance therecognition performance. A two-stage multi-task learning rep-resentation method was also proposed in [340]. After findingan effective subset of training samples and constructing a newdictionary by multi-feature joint sparse representation learningas the first stage, the authors utilized multi-task collaborativerepresentation to perform target images classification basedon the new dictionary in second stage. A multi-level deepfeatures-based multi-task learning algorithm was developed in[347] for SAR-ATR. This architecture employed joint sparserepresentation as the basic classifier and achieved an recog-nition rate of 99.38% on MSTAR under standard operatingconditions (SOCs).A mixed framework based on multimodal, multidiscipline,and data fusion strategy was proposed in [449] for SAR-ATR. An adaptive elastic net optimization method was ap-plied to balance the advantages of l − norm and l − norm optimization on scene SAR imagery by a clustered AlexNetwith sparse coding. The clustered AlexNet with a multiclassSVM classification scheme was proposed to bridge the visual-SAR modality gap. This framework achieved 99.33% and99.86% for the three and ten-class problems on MSTAR,respectively. A SAR and infrared (IR) sensors based multistagefusion stream strategy with dissimilarity regularization usingCNN architecture was developed in [363] to improve the performance of SAR target recognition. In order to make fulluse of phase information of PolSAR images and extract morerobust discriminative features with multidirection, multiscale,and multiresolution properties, a complex Contourlet CNNwas proposed in [376].However, most of DL based SAR-ATR methods presenta limitation that each learning process only handles staticscattering information with prepared SAR image, while miss-ing the space-varying information. To involve space-varyingscattering information to improve the accuracy rate of recog-nition, a novel multi-aspect-aware method was proposed in[326] to learn space-varying scattering information throughthe bidirectional LSTM model. The Gabor filter and three-patch local binary patterns were progressively implemented toextract comprehensive spatial features of multi-aspect space-varying image sequences. After dimensionality reduction withMLP, a bidirectional LSTM learned the multi-aspect featuresto achieve target recognition. This method achieved accuracyrate of 99.9% on MSATR data.To fully exploit the characteristics of continuous SARimaging instead of utilizing single image for recognition,a bidirectional convolution-recurrent network (BCRN) wasdeveloped in [334] for SAR image sequence classification.Spatial features of each image were extracted through DC-NNs without the fully connected layer, and then sequencefeatures were learned by bidirectional LSTM networks toobtain the classification results. In order to exploit the spatialand temporal features contained in the SAR image sequencesimultaneously, a spatial-temporal ensemble convolutional net-work (STEC-Net) was proposed for a sequence SAR targetclassification in [365], which achieved a higher accuracy rate(99.93%) in the MSTAR dataset and exhibited robustness todepression angle, configuration, and version variants. A SARsequence image target recognition network based on two-dimensional (2D) temporal convolution was proposed in [374],including three stages: feature extraction, sequence modelingand classification. To using rotation information of PolSARimage for improving classification performance, the authorsbuilt a convolutional LSTM (ConvLSTM) along a sequence ofpolarization coherent matrices in rotation domain for PolSARimage classification in [375].In order to automatically and precisely extract waterand shadow areas in SAR imagery, the authors proposedmulti-resolution dense encoder and decoder (MRDED) net-work framework in [381], which integrated CNN, ResNet,DenseNet, global convolutional network (GCN), and ConvL-STM. MRDED outperformed by reaching 80.12% in pixelaccuracy (PA) and 73.88% in intersection of union (IoU) forwater, 88% in PA and 77.11% in IoU for shadow, and 95.16%in PA and 90.49% in IoU for background classification,respectively. A feature recalibration network with multi-scalespatial features (FRN-MSF) was built in [382], which achievedhigh accuracy in SAR-based scene classification. FRN wasused to learn multi-scale high-level spatial features of SARimages, which integrated the depthwise separable convolution(DSC), SE-Net block and CNN.In order to make full use of pose angle information andintensity information of SAR data for boosting target recog- OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 21 nition performance, a CNN-based SAR target recognition net-work with pose angle marginalization learning, called SPAM-Net was proposed in [367] that marginalized the conditionalprobabilities of SAR targets over their pose angles to pre-cisely estimate the true class probabilities. A combination ofmulti-source and multi-temporal remote sensing imagery wasproposed in [368] for crops classification, which used CNNand visual geometry group (VGG) to classify crops based onthe different numbers of input bands composed by opticaland SAR data. Deep bimodal autoencoders were proposedfor classification of fusing SAR and multispectral imagesin [369], which was trained to discover both independen-cies of each modality and correlations across the modalities.Combining polarimetric information and spatial information, adual-branch DCNN (dual-DCNN) was proposed to realize theclassification of PolSAR images in [370]. The first branch wasused to extract the polarization features from the 6-channelreal matrix, which are derived from the complex coherencymatrix. The other was utilized to extract the spatial features ofa Pauli RGB (Red Green Blue) image. These extracted featureswere first combined into a fully connected layer sharing thepolarization and spatial property. Then, the softmax classifierwas employed to classify these features. ii) Enhancing Generalization with Limited Labeled DataData augmentation based technology
To eliminate theoverfitting of small dataset in SAR-ATR, a CNN model withfeature extractor and softmax classifier, combining with dataaugmentation technique, was proposed in [310]. An improvedDQN method for PolSAR image classification was proposedin [388], which could generate amounts of valid data byinteracting with the agent using the ε -greedy strategy. A multi-view DL framework was proposed in [325] for SAR-ATRwith limited data, which introduced a unique parallel deepCNN topology to generate multi-view data as inputs of model.The distinct multi-view features were fused in different layersprogressively. A Gabor-DCNN was proposed to overcomethe overfitting problem due to limited data in [399]. Multi-scale and multi-direction-based Gabor features and a DCNNmodel were used for data augmentation and for SAR imagetarget recognition, respectively. A novel adversarial AE wasproposed to improve the orientation generalization ability forSAR-ATR tasks in [400], which learned a code-image-codecyclic network by adversarial training for the purpose ofgenerating new samples at different azimuth angles. A newdual-channel CNN was developed in [403] for PolSAR imageclassification when labeled samples were small, which firstlyused a neighborhood minimum spanning tree to enlarge thelabeled sample set and then extracted spatial features by DC-CNN model.A DNN-based semi-supervised method was proposed in[408] to tackle the PolSAR image classification when labeledsamples was limited. The class probability vectors were usedto evaluate the unlabeled samples to construct an augmentedtraining dataset. The feature augmentation and ensemble learn-ing strategies were proposed in [398] to address the limitedsamples issue in SAR-ATR tasks. The cascaded featuresfrom optimally selected convolutional layers were concate-nated to provide more comprehensive representation for the recognition. The adaboost rotation forest was introduced toreplace the original softmax layer to realize a more accuratelimited sample-based recognition task with cascaded features.In [420], a superpixel restrained DNN-based multiple deci-sions strategy, including nonlocal decision and local decision,was developed to select credible testing samples. The finalclassification map was determined by the deep network, whichwas updated by the extended training set. Fine-grained DNN structure design-based technology
In[300], [309], the authors used convolutional layer to replacefull connection layer and proposed deep memory CNNs(M-Net) to overcome overfitting caused by small samplesdata, which achieved accuracy rate of more than 99% onMSTAR. Aiming to improve the classification performancewith greatly reduced annotation cost, the authors proposed anactive DL approach for minimally-supervised PolSAR imageclassification [401], which integrated active learning and fine-tuning CNN into a principled framework. A microarchitecturecalled CompressUnit-based deeper CNN was proposed in[404]. Compared with the fewest parameters-based networksfor SAR image classification, this architecture was deeperwith only about 10% of parameters. An efficient transferredmax-slice CNN with L2-regularization term was proposed in[409] for SAR-ATR, which could enrich the features andrecognize the targets with superior performance with smallsamples. An asymmetric parallel convolution module wasconstructed in [410] to avoid severe overfitting. In [411], theauthors developed a systematic approach, based on sliding-window classification with compact and adaptive CNNs, toovercome drawbacks of limited labelled data whilst achievingstate-of-the-art performance levels for SAR land use/coverclassification.TL methods are significantly used in DNN design to solvethe problems caused by limited data. A TL-based algorithmwas proposed in [329] to transfer knowledge, learned fromsufficient unlabeled SAR scene images, to labeled SAR targetdata. The proposed CNN architecture consisted of a classi-fication pathway and a reconstruction pathway (i.e., stackedconvolutional auto-encoders), together with a feedback bypassadditionally. A large number of unlabeled SAR scene imageswere used to train the reconstruction pathway at first. Then,these pre-trained convolutional layers were reused to transferknowledge to SAR target classification tasks, combining withreconstruction loss introduced by feedback bypass.TL strategy was used to effectively transfer the priorknowledge of the optical, non-optical, hybrid optical and non-optical domains to the SAR target recognition tasks in [440].The approach of transferring knowledge from electro-opticaldomains to SAR domains was developed in [406] to eliminatethe need for huge labeled data in the SAR classification.This method learned a shared domain-invariant embeddingby cross-domain knowledge transfer pattern. The embeddingwas discriminative for both related electro-optical and SARtasks, while the latent data distributions of both domainsremained similar. Two TL strategies, based on FCN and U-net architecture, were proposed in [422] for high-resolutionPolSAR image classification with only 50 image patches. Thedistinct pretraining datasets were also applied to different
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 22 scenarios. To adapt deep CNN model for PolSAR targetdetection and classification with limited training samples whilekeeping better generalization performance, expert knowledgeof target scattering mechanism interpretation and polarimetricfeature mining were incorporated into CNN to assist the modeltraining and improve the final application performance [416].The semi-supervised and unsupervised learning methods arealso the significant technologies to alleviate the overfittingwith small labeled data. A semi-supervised TL method basedon GAN was presented in [387] to address the insufficientlabeled SAR data. Firstly, A GAN was trained by variousunlabeled samples to learn generic features of SAR images.Subsequently, the learned parameters were readopted to ini-tialize the target network to transfer the generic knowledge tospecific SAR target recognition task. Lastly, the target networkwas fine-tuned by using both the labeled and unlabeled trainingsamples with a semi-supervised loss function. In [389], anunsupervised multi-level domains adaptation method basedon adversarial learning was proposed to solve the problemof time-consuming for multi-band labeled SAR images clas-sification. A semi-supervised recognition method combiningGAN with CNN was proposed in [390]. A dynamic adjustablemulti-discriminator GAN architecture was used to generateunlabeled images together with original labeled images asinputs of CNN. In order to alleviate the time-consumingproblems of obtaining the labels of radar images, a semi-supervised learning method based on the standard DCGANswas presented in [415]. Two discriminators sharing the samegenerators for joint training.To alleviate the burden of manual labeling, a CNN-basedunsupervised domain adaptation model was proposed in [393]to learn the domain-invariant features between SAR imagesand optical aerial images for SAR image retrieving. An unsu-pervised learning method to achieve SAR object classificationwith no labeled data was introduced in [405]. This approachregared object clustering as a recurrent process, in which datasamples were gradually clustered together according to theirsimilarity, and feature representations of them were obtainedsimultaneously. To address the problem of insufficient labelledtraining, an unsupervised DL model was implemented inthe encoding-decoding architecture to learn feature maps atdifferent scale and combine them together to generate featurevectors for SAR object classification in [402].In [391], the authors employed an extension of WassersteinAE as deep generative model for SAR image generation toachieve SAR image target recognition with high accuracy.A novel generative-based DNN framework was proposedin [392] for zero-shot learning of SAR-ATR. A generativedeconvolutional neural network was referred to as a generatorto learn a faithful hierarchical representation of known targets,while automatically constructing a continuous SAR targetfeature space spanned by orientation-invariant features andorientation angle. In [407], the authors proposed a new few-shot SAR-ATR method based on conv-biLSTM prototypicalnetworks. A conv-biLSTM network was trained to map SARimages into a new feature space where it was easy for clas-sification. Then, a classifier based on Euclidean distance wasutilized to obtain the recognition results. A virtual adversarial regularization term was introduced in a neural nonlocal stackedSAEs architecture to regularize the network for keeping thenetwork from being overfitting [413]. A multilayer AE, com-bining with Euclidean distance as a supervised constraint, tobe used in [394] for SAR-ATR tasks with the limited trainingimages.A new deep network in the form of a restricted three-branchdenoising auto-encoder (DAE) was proposed in [395] to takethe full advantage of limited training samples for SAR objectclassification. In this model, a modified triplet restriction,that combined the semi-hard triplet loss with the intra-classdistance penalty, was devised to learn discriminative featureswith a small intra-class divergence and a large inter-classdivergence. In order to solve overfitting problem, the authorsintroduced a dual-input Siamese CNN into the small samplesoriented SAR target recognition in [396]. The recognitionaccuracy rate of this method outperformed the SVM, A-ConvNet, and 18-layers ResNet by 31%, 13%, and 16%,respectively, in the experiment of 15 training samples and 195testing data. A novel method of target classification of SARimagery based on the target pixel grayscale decline with agraph CNN was introduced in [397], which transformed theraw SAR image from Euclidean data to graph-structured databy a graph structure and these transformed data were as theinputs of graph CNN model. To balance the anti-overfittingand features extraction abilities with small training samplesfor SAR targets images classification, the authors proposeda novel hinge loss (HL)-based CAE semi-greedy network in[412], i.e., CAE-HL-CNN. Compared with existing state-of-the-art network, the CAE-HL-CNN had best performances inclassification accuracy and computation costs with the SOCand EOC MSTAR datasets. iii) Improving Robustness of Recognition Algorithms
The speckle noise, clutter, scale-variance of inputs, andadversarial samples can severely cause unstability of DNNalgorithm in SAR-ATR. Therefore, the robustness improve-ment of DNN algorithms is very vital. A new multi-viewsparse representation classification algorithm based on jointsupervised dictionary and classifier learning was developed in[336] for SAR image classification. During training peocess,classification error was back propagated to the dictionarylearning procedure to optimize dictionary atoms. In this way,the representation capability of the sparse model was en-hanced. This new architecture was more robust for depres-sion variation, configuration variants, view number, dictionarysize, and noise corruption, compared to other state-of-the-artmethods, such as SVM.SAR-ATR is performed on either global or local featuresof acquired SAR images. The global features can be easilyextracted and classified with high efficiency. However, theylack of reasoning capability thus can hardly work well underthe EOCs. The local features are usually more difficult toextract and classify, but they can provide reasoning capabilityfor target recognition. To make full use of global and local fea-tures of SAR-ATR at the EOCs, a hierarchical fusion schemeof the global and local features was proposed in [330] tojointly achieve high efficiency and robustness in ATR system.The global random projection features can be extracted and
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 23 classified by sparse representation-based classification mthodeffectively. The physical related local descriptors, i.e., ASCs,were employed for local reasoning to handle various EOCs likenoise corruption, resolution variance, and partial occlusion.To improve robustness of model for noise and invariance ofmodels, a multiple feature-based CNN model was employedin [332] to recognize the SAR image in an end-to-end learningway. The strong features more effected by noise and smoothedfeatures less influenced were aggregated into a single columnvector to build complementary relationships for recognitionby a full connection network. As for target rotation behaviorrecognition, a rotation awareness based self-supervised DLmodel was proposed in [335]. This model suggested thatmore attention should be paid on rotation-equivariant andlabel-invariant features. To explore the property of translation-invariance of CNNs, the authors verified that ResNet couldachieve translation-invariance with aligned SAR images [311],even ResNet do not adopt data augmentation. A scale-invariantframework based on CNN was proposed in [427] to improvethe robustness of model with respect to scale and resolutionvariations in dataset. This architecture developed an uniformrepresentation method to enlarge the feature space for thevariants of data by concatenating scale-variant and scale-invariant features.Luminance information of SAR images was used to formthe target’s profile in [417] to significantly reduce the influenceof speckle noise on CNN model. A scale transformationlayer was embedded in deep convolutional autoencoder modelto reduce the influence of noise in [418]. To restrain theinfluence of speckle noise and enhance the locally invariantand robustness of the encoding representation, the operationsof contractive restriction and graph-cut-based spatial regular-ization in DSCNN were adopted in [419]. A SRDNN-basedSAE was proposed to capture superpixel correlative featuresto reduce speckle noises in [420]. A speckle-noise-invariantCNN was developed in [421], which employed regularizationterm to improve Lee sigma filter performance, i.e., minimizingfeature variations caused by speckle noise.A TL-based top-2 smooth loss function with cost-sensitiveparameters was introduced to tackle the problems of labelnoise and imbalanced classes in [422]. A CNN-based recog-nition method of synthetic SAR dataset with complex back-ground was proposed in [423]. As for noise and signal phaseerrors, the authors proposed a advanced DL based adversarialtraining method to mitigate these influence in [424]. A point-wise discriminative auto-encoder was proposed in [425] toextract noise and clutter robust features from the target areaof SAR images. In order to alleviate the speckle influence onthe scattering measurements of individual pixels in PolSARimages, local spatial information was introduced into stackedsparse autoencoder to learn the deep spatial sparse featuresautomatically in [426].Moreover, The DL-based SAR target recognition algorithmsare potentially vulnerable to adversarial examples [428]. In[424], the authors involved a adversarial training technologyto ensure the robustness of DL algorithm under the attacksof adversarial samples. HySARNet, as a hybrid ML model,was proposed in [429] to determine the robustness of model when faced variations in graze angle, resolution, and additivenoise in SAR-ATR tasks. A wavelet kernel sparse deep codingnetwork under unbalanced dataset was proposed in [430] forunbalanced PolSAR classification.The issue of different characters of heterogeneous SARimages will lead to poor performances of TL algorithm inSAR image classification. To address this problem, a semi-supervised model named as deep joint distribution adaptationnetworks was proposed in [431] for TL model, which learningfrom a source SAR images to similar target SAR images.In order to increase the stability of GANs model training inSAR targets recognition, the authors proposed a new semi-supervised GANs with multiple generators and a classifier in[414]. Multiple generators were employed to keep stability oftraining. iv) Promoting the Real-Time or Reducing ComputationCosts
A CNN-based framework consisted of SqueezeNet networkand a modified wide residual network was developed in [298]to build real-time damage mapping for classifying differentdamaged regions on the SAR image. A direct ATR methodwas employed in [346] for large-scene SAR-ATR task, whichdirectly recognized targets on large-scene SAR images byencapsulating all of the computation in a single DCNN.Experiments on MSTAR and large-scene SAR images (withresolution 1478 * 1784) showed this model outperformedother methods, such as CFAR+SVM, region-based CNN, andYOLOv2 [466]. The PCANet was employed in [348] for SAR-ATR to achieve more than 99% accuracy rate on MSTAR. A-convNet was proposed in [259] to achieve an average accuracyrate of 99.1% on MSTAR. A novel stacked deep convolutionalhighway unit network was proposed in [323] for SAR imageryclassification, which achieved accuracy rate of 99% with allMSTAR data, and still reached 94.97% when the training datawas reduced to 30%.The complex multi-view processing of images, however,can cause huge computation costs for multi-view learningmethod. To address this problem, a optimal target viewpointsselection based multi-view ATR algorithm was developed in[328]. This algorithm used two-channel CNNs as multi-viewclassifiers, which was based on ensemble learning [51]. Adirect graph structure-based single source shortest path searchalgorithm was also adopted to represent the tradeoff betweenthe recognition performance and flight distance of SAR plat-form. A heterogeneous CNN-based ensemble learning methodwas employed in [447] to construct noncomplete connectionscheme and multiple filters stacked.A lightweight CNN model was designed in [341] to recog-nize the SAR images. The channel attention by-pass and spa-tial attention by-pass were introduced to enhance the featureextraction ability. Depthwise separable convolution was usedto reduce the computation costs and heighten the recognitionefficiency. In addition, a new weighted distance measured lossfunction was introduced to weaken the adverse effects of dataimbalance on accuracy rate of minority class. This architecturehas better performance than ResNet, A-ConvNet [259], [342].A one-layer based novel incremental Wishart broad learningsystem was specifically designed in [432] to achieve PolSAR
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 24 image classification, which could effectively transfer essentialWishart distribution and other types of polarimetric decom-position and spatial features to establish feature map andenhancement nodes in just one layer without DL structures.Therefore, the training consumption could be decreased sig-nificantly. Similarly, a superpixel-driven optimized Wishartnetwork was introduced in [433] for fast PolSAR imageclassification. In [434], the authors applied some tricks (suchas BN, drop-out strategy) and concatenated ReLU to reducecomputation cost of DL algorithm. A spatial-anchor graphbased fast semi-supervised classification algorithm for PolSARimage was introduced in [435].In [436], the authors proposed a novel method based ontarget pixel grayscale decline by a graph representation net-work to accelerate the training time and achieve classificationaccuracy rate of 100%. In order to speed up computationand improve classification accuracy, a classification methodof full-polarization SAR images, based on DL with shallowfeatures, was proposed in [437]. Aiming to solve the problemsof energy consumption, so as to deploy the DL model on em-bedded devices conveniently and train the model in real-time,a custom AI streaming architecture was employed in [438]for SAR maritime target detection. A more flexible structureas the new implementation of CNN classifier was proposedin [439], which had less free parameters to reduce trainingtime. An atrous-inception module-based lightweight CNN wasproposed in [440], which combined both atrous convolutionand inception module to obtain rich global receptive fields,while strictly controlling the parameter amount and realiz-ing lightweight network architecture. In [441], apache sparkclustering framework was presented for classification of high-speed denoised SAR image patches. An asymmetric parallelconvolution module was constructed in [442] to alleviate thecomputation cost. In order to alleviate the trade-off betweenreal-time and high performance, the authors proposed a semi-random DNN to exploit random fixed weights for real-timetraining with comparable accuracy of general CNNs in [443].To tackle the issues of low memory resources and lowcalculation speed in SAR sensors, the authors proposed a amicro CNN for real-time SAR recognition system in [444],which only had two layers, compressed from a 18-layerDCNN by a novel knowledge distillation algorithm, i.e.,gradual distillation. Compared with the DCNN, the memoryfootprint of the proposed model was compressed 177 times,and the computation costs was 12.8 times less. In order todeploy a real-time SAR platform, three strategies of net-work compression and acceleration were developed in [445]to decrease computing and memory resource dependencieswhile maintaining a competitive accuracy. Firstly, weight-based network pruning and adaptive architecture squeezingmethod were proposed to reduce the consumption of storageand computation time of inference and training process of DLmodel. Then weight quantization and coding were employedto compress the network storage space. In addition, a fastapproach for pruning convolutional layers was proposed toreduce the number of multiplication by exploiting the sparsityof the inputs and weights.At present, most of neutral network-based classification methods need to expand the dataset by data augmentationtechnology or design the light-weighted network model toimprove their classification performance. However, optimaltraining and generalization are two main challenges for DNNmodel. Instead of DNN model, a novel deep forest model wasconstructed in [446] by multi-grained cascade forest (gcForest)to classify 10-class targets on MSTAR. This was the firstattempt to classify SAR targets using the non-neural networkmodel. Compared with DNN-based methods, gcForest hadbetter performances in calculation scale, training time, andinterpretability.
4) Ship Targets Detection based on SAR Images .In section 3) we make a comprehensive survey on SAR-ATR based on DL algorithm. From the overview in publishedliteratures, the SAR-ATR is a very important research domainwidely involved in military and civil applications. The SAR-based ship targets detection (STD), one of the importantresearch aspects in maritime surveillance (such as marinetransportation safety), is an another significant research direc-tion for SAR image processing. Of course, optical imagery-based ship detection and classification is also a hot researchdirection, please refer to [35]. The SAR images of the STDusually have a large scale, which contains many different scaleship targets. The goal of STD is detection and recognition ofeach target on the SAR image.Traditional STD approaches include constant false alarmrates (CFAR) based on the distributions of sea clutter [453],[454], extracted features manually based on ML algorithm[455]–[458], dictionary-based sparse representation, SVM,template matching, K-NN, Bayes, saliency object detectionmodels. Traditional methods, however, intensively depend onthe statistics modeling and the experts’ feature extractionability, which degrades the detection performances of SARimagery to some extend.In recent years, DL-based methods have produced manygreat achievements in objects detection domain. These DLalgorithms can be roughly categorized into two classes: two-stage methods and one-stage methods. The former firstly gen-erates positive region proposals to discard the most of negativesamples, then performs the candidate regions classification,such as region convolutional neural networks (R-CNN) [459],fast R-CNN [460], faster R-CNN [461], mask R-CNN [462],cascade R-CNN [463], feature pyramid networks (FPNs)[464]. The latter directly detects the objects by obtainingobjects’ coordinate values and the class probability, whichconsiders both accuracy and computation costs, such as YouOnly Look Once (YOLOs: v1-v4, poly-v3) [465]–[469], singleshot multiBox detector (SSD) [470], RetinaNet [471]. Thetwo-stage methods have higher accuracy, but slower trainingthan one-stage methods.Nowadays, the SAR researchers have successfully appliedDL algorithms in STD. Some challenges, however, haveoccurred in this domain even though applied DL algorithms,which mainly focus on three aspects: (i) ships often have alarge aspect ratio and arbitrary directionality in SAR images.Traditional detection algorithms can unconsciously cause re-dundant detections, which make it difficult to accurately locatethe target in complex scenes (such as background interference,
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 25 clutter, inshore and outshore scenes, e.g., Chinese Gaofen-3 (GF-3) Imagery has 86 scenes [480]); (ii) ships in portsare often densely arranged, and the effective identification ofthese ships is complicated; and (iii) ships in SAR imageshave various scales due to the multi-resolution imaging modesemployed in SAR (such as GF-3 Imagery has four resolutions,i.e., 3 m, 5 m, 8 m, and 10 m [480]) and various ship shapes,which pose a considerable challenge for STD. In this section,we will do a comprehensive survey on DL-based STD, whichmainly focuses on solving these challenges aforementioned.In [472], faster R-CNN architecture [461] was investigatedin STD. A new dataset and four strategies (feature fusion,transfer learning, hard negative mining, and other implemen-tation details) were proposed to achieved better accuracy andless test computation costs than the standard faster R-CNNalgorithm. A densely connected multi-scale neural networkbased on faster R-CNN framework was proposed in [481]to solve the multi-scale and multi-scene STD problems. In[491], the authors proposed a ship detection and segmentationmethod based on an improved mask R-CNN model [462]. Thismethod could accurately detect and segment ships at the pixellevel. In [480], RetinaNet [471] model was used as the objectdetector, to automatically determine features representationlearning effectively for multi-scale ships detection.However, the common target detection models originatefrom the optical image detection tasks in CV, which maybedegrade their performances when applied to STD more or less,because of special imaging principles of SAR images. Manynew algorithms have been proposed to specially address thechallenges in STD. These new ideas remain depending onbasic targets detection models, such as FPNs [464], R-CNN[460]. i) Improving Accurately Location of Ship Targets
As for the first problem, i.e., it is difficult to accuratelylocate the targets in complex scenes. RetinaNet was appliedto [480] to alleviate the limitation of highly depending onthe statistical models of sea clutter in STD, which achievedmore than a mean average precision (MAP) of 96%, and couldefficiently detect multi-scale ships with high effectivenessin GF-3 SAR images. A new land masking strategy basedon the statistical model of sea clutter and neural networkmodel was employed in [477] to detect ships in GF-3 SARimages. The fully convolutional network (FCN) was appliedto separate the sea area from the land. Then, choosing theprobability distribution model of CFAR detector based on atradeoff between the sea clutter modeling accuracy and thecomputational complexity. In addition, truncated statistic anditerative censoring scheme were used to better implementCFAR detection for boosting the performance of detector.Due to the multi-resolution imaging mode and complex back-ground, multi-level sparse optimization method of SAR imagewas studied in [474] to handle clutters and sidelobes, so as toextract discriminative features of SAR images. A segmentationmethod based on a U-Net was developed in [473] to addressthe problems of false alarms caused by ocean clutter. Thisalgorithm was designed specifically for pixel-wise STD fromcompact polarimetric SAR images. A novel object detectionnetwork was employed in [476], [485] to extract contextual features of images. This model also used attention mechanismto rule out false alarms in complex scenarios. A new trainingstrategy was adopted in [481] to reduce the weights of easyexamples in the loss function, so that more attention focusedon the hard examples in training process to reduce false alarm.Two parallel sub-channels based multi-feature learning frame-work was proposed in [482], including DL-based extractedfeatures and hand-crafted features. Two sub-channels featureswere concatenated to extract fused deep features to achievehigh performance. ii) Accurately Detection of Densely Arranged Ships
As for second problem, it is difficult to detect denselyarranged ships. Non-maximum suppression (NMS) methodwas widely used to address this issue. A soft-NMS method wasintroduced into the detection network model in [485], [492]to reduce the number of missed detections of ship targets inthe presence of severe overlap for improving the detectionperformance of the dense ships. In addition, the modifiedrotation NMS was developed in [488] to solve the problemof the large overlap ratio of the detection box. iii) Solving the Problems of Multi-scale Variations
More importantly, it is very vital to design a optimal solutionto solve the problems of multi-scale variations in design ofSTD algorithms. A FPN was used in [480] to extract multi-scale features for both ship location and classification, andfocal loss was also used to address the class imbalance toincrease the importance of the hard examples during trainingprocess. A densely connected multi-scale neural network basedon faster R-CNN was proposed in [481] to densely connectone feature map to each other feature maps from top to down.In this way, the positive proposals were generated from eachfused feature map based on multi-scale SAR images in multi-scene. Similarly, combining with densely connecting convo-lutional block attention module, a dense attention pyramidnetwork was developed in [487], [490] to concatenate featuremaps from top to bottom of the pyramid network. In this way,sufficient resolution and semantic information features wereextracted. In addition, convolutional block attention modulerefined concatenated feature maps to fuse highlight salientfeatures with global unblurred features of multi-scale ships,and the fused features were as the inputs of detection networkto accurately obtain the final detection results.To address the diverse scales of ship targets, a loss functionincorporated the generalized intersection over union (GIoU)loss to reduce the scale sensitivity of the network [485].In [486], a new bi-directional feature fusion module wasincorporated in a lightweight feature optimizing network toenhance the salient features representation of both low andhigh features representation layers. Aiming to fast achieve po-sitioning rotation detection, the authors proposed a multiscaleadaptive recalibration network in [488] to detect multiscaleand arbitrarily oriented ships in complex scenarios. The re-calibration of the extracted multiscale features improved thesensitivity of the network to the target angle through globalinformation. In particular, a pyramid anchor and a loss functionwere designed to match the rotated target to accelerate therotation detection.To eliminate the missing detection of small-sized ships
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 26 targets in SAR imagery, a contextual region convolutionalhierarchical neural network with multilayer fusion strategywas designed in [476], which consisted of a high resolutionRPN and an object detection network to extract contextualfeatures. This framework fused the deep contextual semanticfeatures and shallow high-resolution features to improve thedetection performance for small-sized ships. A novel splitconvolution block was used in [474] to enhance the featurerepresentation of small targets, which divided the SAR imagesinto smaller sub-images as the inputs of the network. Also,a spatial attention block was embedded in FPN to reduce theloss of spatial information during the dimensionality reductionprocess.Based on TL method, a pre-trained YOLOv2 model [466]was applied to STD in [483]. The experiments on threedifferent datasets showed the effectiveness of the pre-trainedYOLOv2. TL strategy was also used to train the detectionmodel in [478] due to the limited number of datasets. In-stead of a single feature map, a scale-transferrable pyramidnetwork was employed in [489] for multi-scale detection. Alatent connection based FPN was constructed to inject moresemantic information into feature maps with high resolution,and densely connected each feature maps from top to down byusing scale-transfer layer. Therefore, the dense scale-transferconnection could expand the resolution of feature maps andexplicitly explore valuable information contained in channels.A scale transfer module was also used in [484] to connect withseveral feature maps to extract multiscale features for STD. Inaddition, RoIAlign was adapted to calibrate the accuracy ofthe bounding boxes, and the context features were employed toassist the detection of complex targets in detection subnetwork.Nowadays, the existing methods of SAR STD mainly de-pend on low-resolution representations obtained by classi-fication networks or recover high-resolution representationsfrom low-resolution representations in SAR images. Thesemethods, however, are difficult to obtain accurate predictionresults in spatial accuracy of region-level. Based on a high-resolution STD network, a novel framework was proposed in[492] for high-resolution SAR imagery ships detection. Thisarchitecture adopted a novel high-resolution FPN connectingwith several high-to-low resolution subnetworks in parallel, tomake full advantage of the high-resolution feature maps andlow-resolution convolutions to maintain high resolution STD.In addition, soft-NMS was also used to improve the detectionperformance of the dense ships and the Microsoft COCOevaluation metrics was introduced for performance evaluation.Most of STD algorithms are focus on detection accuracy.Detection speed, however, is usually neglected. The speed ofSAR STD is extraordinarily important, especially in real-timemaritime rescue and emergency military decision-making. Toimprove the detection speed, a pyramid anchor and a lossfunction were designed in [488] to match the rotated targetsto speed up the arbitrary ships rotation detection. A novelgrid CNN was developed in [493] for high-speed STD, whichmainly consisted of a backbone CNN and a detection CNN. In-spired by the idea of YOLO algorithm, this method improvedthe detection speed by meshing the input images and usingthe depthwise separable convolutions. The experiments results on SSDD dataset and two SAR images from RadarSat-1 andGaofen-3 showed that the detection speed of this model wasfaster than the other existing methods, such as faster R-CNN,SSD, and YOLO under the same computing resource, andthe detection accuracy was kept within an acceptable range.To infer a large volume of SAR images with high detectionaccuracy and relatively high speed, SSD was adopted in [478]to address STD in complex backgrounds. TL strategy was alsoadopted to train the detection model.In sections 3) and 4), we make a comprehensive surveyof SAR-ATR and STD based on SAR imagery. In addition,SAR imagery segmentation is also researched. Targets seg-mentation tries to separate the target from the background thuseliminating the interference of background noises or clutters.However, it may also discard a part of the target characteristicsand target shadows during the segmentation process, whichalso contains discriminative information for target recogni-tion. Then the tradeoff between interference elimination anddiscriminability loss will degrade target recognition to someextent [496]. Therefore, the comprehensive evaluation for theeffectiveness of segmentation on target recognition is veryimportant. A novel architecture for SAR segmentation basedon convolutional wavelet neural network (CWNN) and MRF[495] were proposed in [494], which could suppress the noiseand keep the structures of the learned features complement.In addition, a ship detection and segmentation method basedon an improved mask R-CNN model was developed in [491],which could accurately detect and segment ships at the pixellevel. To allow lower layers features to be more effectivelyutilized at the top layer, a channel-wise and spatial attentionmechanisms based bottom-up structure was added to FPNstructure of mask R-CNN, so as to shorten the paths betweenlower layers and the topmost layer. The experiments resultsshowed that the MAPs of detection and segmentation increasedfrom 70.6% and 62.0% to 76.1% and 65.8%, respectively.
B. ISAR Images Processing
1) ISAR Imaging .To address the problem of low-resolution (LR) ISAR imag-ing, the authors employed deep ResNet as an end-to-endframework to directly learn the mapping between the inputLR images and the output high-resolution (HR) images withrespect to the point spread function (PSF) in [497]. An amountof multiplicative noise or clutter may be present in real-worldISAR measurement scenarios. The current linear imagingmethods are not generally well suitable to alleviate the effectsof noise, such as MUSIC, compressive sensing (CS). Sincethese algorithms rely on phase information significantly whichcan be heavily distorted or randomized under the imagingprocess. The authors introduced CNNs model to deal withthis issue in [498]. In order to exploit a real-time ISARimaging algorithm, the authors proposed an efficient sparseaperture ISAR autofocusing algorithm in [499], which adopteddivided simpler subproblems by alternating direction methodof multipliers and auxiliary variable to alleviate the complexcomputation of ISAR imaging used sparse Bayesian learning(SBL) method. This method achieved 20-30 times faster thanthe SBL-based approach.
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 27
To address the problem of basis mismatch of CS basedHR ISAR imaging of fast rotating targets, a pattern-coupledsparse Bayesian learning method for multiple measurementvectors, i.e., the PC-MSBL algorithm, was proposed in [500].A multi-channel pattern-coupled hierarchical Gaussian priorwas introduced to model the pattern dependencies among theneighboring range cells and correct the migration throughrange cells problem. The expectation-maximization (EM) al-gorithm was used to infer the maximum a posterior estimateof the hyperparameters. To tackle the issue of destroyedcoherence between the undersampled pulses caused by sparseaperture radar echoes, the authors proposed a novel BayesianISAR autofocusing and scaling algorithm for sparse aperturein [501].
2) ISAR Targets Detection and Recognition .In order to tackle the challenges of ISAR objects detection, afast and efficient weakly semi-supervised method, called deepISAR object detection (DIOD), was proposed in [502], whichwas based on advanced region proposal networks (ARPNs)and weakly semi-supervised deep joint sparse learning. Thisframework used i) ARPN to generate high-level region propos-als and localize potential ISAR objects robustly and accuratelyin minimal time, ii) a convenient and efficient weakly semi-supervised training method was proposed to solve the problemof small annotated training data, and iii) a novel sharable-individual mechanism and a relational-regularized joint sparselearning strategy were introduced to further improve the ac-curacy and speed of the whole system. Similarly, the authorsproposed a novel DIOD method, which was based on fullyconvolutional region candidate networks and DCNNs in [503].A TL-based novel method of multiple heterogeneous pre-trained DCNN (P-DCNN) ensemble with stacking algorithmwas firstly proposed in [504], which could realize automaticrecognition of space targets in ISAR images with high ac-curacy under the condition of the small samples. The stack-ing algorithm was used to realize the ensemble of multipleheterogeneous P-DCNNs, which effectively overcame weakrobustness and difficulty in classification accuracy existing ina single weights fine-tuned P-DCNN. A semantic knowledgebased deep relation graph learning was proposed in [505]for real-world ISAR object recognition and relation discovery.Dilated deformable CNN was introduced to greatly improvesampling and transformation ability of CNN, and increasethe output resolutions of feature maps significantly. Deepgraph attribute-association learning method was proposed toobtain semantic knowledge to exploit inter-modal relationshipsamong features, attributes, and classes. Multi-scale relational-regularized convolutional sparse learning was employed tofurther improve the accuracy and speed of the whole system.In addition, CNNs and CAEs were also used to classify ISARobjects in [506].Three ML algorithms were introduced in [507] for ISAR tar-gets classification, i.e., DT, Bayes, and SVM. A SAE learningalgorithm was employed in [508] to solve the classificationissue of non-cooperative airplane targets with ISAR images.
C. HRRP-based Automatic Target Recognition
With the advantages of easily acquisition, processing andabundant target feature information, unidimensional high res-olution range profile (HRRP) is a specially concern researchdirection of ATR. HRRP is the projection of target echo scattervectors in the direction of radar sight line, at the conditionof big transmitted signal bandwidth and big target shape.The HRRP-ATR research domain mainly concerns solvingthree aspect problems: noise robustness, discriminative andinformative features extraction, and optimal classifier design.In practice, the first two problems are usually tackled simul-taneously.There are three stages for HRRP-ATR: image preprocessing[515], feature extraction and classifier design respectively.Image preprocessing mainly includes denoising [512], [514],[521], [522] and alleviates sensitivity problems: gesture, trans-lation, and amplitude [513], [514], [519], [528]. Feature ex-traction process extracts low dimensional inherent featuresfrom preprocessed HRRP, which are easily identifiable forHRRP of the target, including PCA, expert-based featureengineering. A fine classifier is used to achieve ATR tasks,such as SVM, DNNs.Three stages are not rigorously operated sequentially. Somealgorithms can achieve multi-operation simultaneously, e.g.,PCA has denoise and dimensionality reduction functions. TakeDNNs as an example, the DNNs are end-to-end learningarchitectures, operating the feature extraction and classificationsimultaneously [519], [528], [529].
1) Feature Extraction .Probabilistic principal component analysis (PPCA) modelwas proposed in [509], [510] for noise robust feature ex-traction, which provided prior information for robust featuresfrom statistic modeling perspective. In [511], the authorsadopted Bernoulli-Beta prior to learn the needed atoms todetermine relatedness between frames of training data. Afeature extraction dictionary was used to extract the localand global features of target’s HRRP [512], [523] for multi-feature joint learning method based on sparse representationand low-rank representation. Support vector data descriptionwas developed in [513] to extract non-linear boundary ofdataset as classification features. In addition, orthogonal maxi-mum margin projection subspace (OMMPS) was employed in[514] for HRRP’s feature extraction to reduce redundancy. Toimprove recognition performance, multiple kernel projectionsubspace fusion method was introduced in [514], [516] forfeature extraction of HRRP, this method can guarantee theintegrity of target information and robustness.As for dealing with the challenge of noncooperative targetrecognition with imbalanced training datasets, t-distributedstochastic neighbor embedding (t-SNE) and synthetic sam-pling for data preprocessing were utilized in [517] to provide awell segmented and balanced HRRP dataset. Scatter matchingalgorithm was proposed in [521], [522] for dominant scattersfeatures extraction of HRRP with noise robustness. Multi-scale fusion sparsity preserving projections approach was alsoproposed in [524] to construct multi-scale fusion featuresin each scale and their sparse reconstructive relationship,
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 28 which contained more discriminative information. To exploitpotential features of HRRP, scale space theory based featureextraction method was employed in [525], which extendedfrom single scale to multiple scales.
2) Classifier Design .The classifiers of HRRP-based classification mainly includeSVMs [524], [525], DBNs [517], [518], SAEs [520], [532],and RNNs [528]. The learning strategies of classifier includemultitask learning [511], multi-feature learning [512], andmulti-scale learning [524]–[526].PPCA based dictionary learning method was proposed in[510] for HRRP recognition. OMMPS is used to maximize themargin of inter-class by increasing the scatter distance of inter-class and reducing the scatter distance of intra-class simultane-ously [514], to improve the recognition performance. T-SNEbased discriminant DBN was proposed in [517] as an efficientrecognition framework with imbalanced HRRP data, whichnot only made full use of dataset inherent structure amongHRRP samples for segmentation, but also utilized high-levelfeatures for recognition. Moreover, the model shared latentinformation of HRRP data globally, which could enhance theability of modeling the aspect sectors with few HRRP data.In order to reduce preprocessing works, discriminative in-finite RBM (Dis-iRBM) was proposed in [518] as an end-to-end adaptive feature learning model to recognize HRRP data.Concatenated DNN was used in [519] for HRRP recognition.Multi-evidence fusion strategy was also adopted for recogni-tion of multiple samples to improve performance.Although the deep structure has high accuracy, it is hardto achieve the performance of good generalization and fastlearning. In [520], the authors combined SAE with regularizedELM to recognize HRRP data with a fast learning speedand better generalization performance. SVM was employed toverify the classification performance of features extracted byMSFSPP and related feature extraction methods in [524]. SVMand three nearest neighbor classifiers demonstrated that the ap-plication of scale-space theory in multi-scale feature extractioncould effectively enhance the classification performance [525].A TL-based feature pyramid fusion lightweight CNN modelwas proposed in [526] to conduct multi-scale representationof HRRP target recognition with small samples at low SNRscenario. Reconstructive and discriminative dictionary learn-ing based on sparse representation classification criteria wasdeveloped in [527], which incorporated the reconstructive anddiscriminative powers of atoms during the update of atoms.This algorithm was more robust to the variation of target aspectand noise effect.To extract fine discriminative and informative features ofHRRP, target-aware recurrent attentional network (TARAN)was used in [528] to make use of temporal dependence andfind the informative areas in HRRP. This network utilizedRNN to explore the sequential relationship between the rangecells within a HRRP sample, and employed the attentionmechanism to weight up each time step in the hidden state,so as to discover the target area. To extract high dimensionalfeatures and generally contain more target inherent character-istics, discriminant sparse deep AE framework was proposedin [529] to classify HRRPs with small data samples. This framework was inspired by multitask learning and trained bythe radar HRRP samples to share inherent structure patternsamong the targets. In [532], the authors built stacked correctiveAE to recognize HRRP, which employed the average profileof each HRRP frame as the correction term.Considering the noise robust recognition of noncooperativetargets, Gaussian kernel and Morlet wavelet kernel were com-bined in [530] to form a multiscale kernel sparse coding-basedclassifier to recognize radar HRRP, which had comparableperformance with well-studied template based methods, suchas SVM, sparse coding-based classifiers (SCCs) and kernelSCCs. To classify the FFT-magnitude features of complexHRRP, least square support vector data description classifierwas developed in [531] to classify HRRP with low compu-tational complexity and overcame the shortcoming of poorcapacity of variable targets in support vector data description.
D. Micro-doppler Signature Recognition
Micro-doppler (MD) technique aims to extract the micro-motions of subjects, that may be unique to a particular subjectclass or activity, to distinguish probable false alarms from realdetections or to increase the valuable information extractedfrom the sensor. Using the available MD returns from sensorfor recognition can significantly reduce the false alarm rate,thereby improving the utility of the sensor system [549].Radar MD signatures, derived from these motions, illustratethe potential ability of the joint time-frequency analysis forexploiting kinetic and dynamic properties of objects [550],such as drones [551]–[553], unmanned aerial vehicle (UAV)[551], human motion [555], deceptive jamming [554].The MD-based classification and recognition of object’spostures and activities has widely absorbed research concernsin the past few years, such as human detection and activityclassification [533], human gesture [539] and gait recognition[537], [538], UAV detection [551]. This section will review theachievements of ML-based radar MD signature processing intarget classification and recognition.Similar to HRRP-ATR, feature extraction and classifierdesign are mainly stages for MD signature based recognitiontasks. The features of targets’ activities are extracted from theradar MD spectrogram, such as single vector decomposition(SVD) vectors of raw data [536], [538]. The optimal classifierdesign is based on ML models, such as SVM [533], ANN[534], CNN [535], [539], [541], [548], CAE [541], [543],[545], RNN [547].A novel robust MD signal representation method based onboth magnitude and phase information of the first Fouriertransform was proposed in [540] for UAV detection, i.e.,2D regularized complex-log-Fourier transform and an object-oriented dimensionality reduction technique-subspace reliabil-ity analysis. The latent space representation was extracted andinterpreted in [541] from 2D CAEs and t-SNE, respectively.In addition, CAE architecture was employed in [545] for MDfeature extraction. Three features extraction algorithms wereproposed in [546], spectrogram frequency profile (SFP) algo-rithm, cadence velocity diagram frequency profile (CVDFP)algorithm, and SFP-CVDFP-PCA algorithm, respectively.
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 29
As for classifier design, the SVM and ANN classifier weredeveloped in [533], [534] for seven-class human activities clas-sification based on six features extracted from radar dopplerspectrogram, respectively. These features included running,walking, walking while holding a stick, crawling, boxing whilemoving forward, boxing while standing in place, and sittingstill. Compared to SVM classifier, a 3-layer AE structure wasproposed in [537], which achieved a accuracy rate of 89%,17% improvement compared to the benchmark SVM with 127pre-defined features. SVM was also used in [546] based onmulti-feature integration.Deep CNN structure was proposed in [535] for humandetection and activities classification, jointly learned the neces-sary features and classification boundaries. MD-based humanhand gestures recognition achieved accuracy rate of 98% usingCNN in [539]. A 50-layer ResNet was trained to identify thewalking subject based on the 2D signature [541]. TL-basedDNN model was also used to classify MD signatures of gaitmotion in [542].To seek the efficient MD analysis algorithm to distinguishthe gaits of subjects, even the MD signatures of those gaits arenot visually distinguishable, a 3-layer deep CAE was proposedin [543], which utilized unsupervised pre-training to initializethe weights in the subsequent convolutional layers, and yieldeda correct classification rate of 94.2%, 17.3% improvementover the SVM. These MD signatures ere were measured fromthe 12 different human indoors activities using a 4 GHzcontinuous wave radar. CAE is more efficient than other deepmodels, such as CNN, AE, and traditional classifiers, such asSVM, RF and Xgboost. To compare the efficiency of ANNinitialization technologies in classification of MD signals,an unsupervised TL-based pretraining method was appliedto CAE [544]. VGGNet and GoogleNet were employed toclassify human activities with small training samples. In orderto address the measurements of a variable observation time andtransition between classes over time, a sequence-to-sequenceclassification method, i.e., RNN with LSTM architecture, wasdeveloped in [547]. In addition, to make full use of timeand frequency domain features of MD signatures, mergingtime and frequency-cadence-velocity diagram was proposedin [548] for drone classification with GoogleNet.
E. Range-doppler Image Processing
Range-doppler (RD) images is also used for classificationand recognition of target’s motions. A RD image containsinformation of range units and doppler features. In linearfrequency modulation continuous wave (LFMCW) radar, theRD imaging process is as the following: firstly removingthe slope of echo signal, then obtaining the radical rangeinformation by FFT of fast time domain signal, after that,acquiring the energy distribution of doppler domain by FFTof slow time direction in the same range unit.Two different classification architectures based on SAEwere developed in [556] for human fall detection and classifi-cation, which used RD images and MD images as the inputsof the cascade and parallel connection models, respectively.Firstly, RD images and MD images were as the inputs of initialSAEs to extract identifiable features, respectively. Then, the extracted features were fused as the inputs of a final SAE tofinish classification task. The results of experiment showedthat the detection probabilities were 89.4% and 84.1% forcascade and parallel detection architecture on same dataset,respectively.Combining with convolutional and memory functions, anend-to-end learning architecture based on CNN and LSTMwas developed in [557] for 11 kinds of dynamic gesturesrecognition. The RD images of gestures at a time point wereas the inputs of CNN and then the RD images sequencesat different time points were as inputs of LSTM to finishrecognition. This novel recognition model achieved averageaccuracy rate of 87% on a 11 kinds of dynamic gestures dataand generalized well across 10 users.A novel detection method was developed in [558] for re-motely identifying a potential active shooter with a concealedrifle/shotgun based on radar MD and RD signatures analysis.Special features were extracted and applied for detecting peo-ple with suspicious behaviors. ANN model was also adoptedfor the classification of activities, and achieved an accuracyrate of 99.21% in distinguishing human subjects carrying aconcealed rifle from other similar activities.V. A
NTI - JAMMING AND I NTERFERENCE M ITIGATION
This section will review the radar anti-jamming and interfer-ence mitigation technologies in ML-related RSP domain, in-cluding jamming or interference classification and recognitionand anti-jamming and interference mitigation strategies. Theexamples of 2D time-frequency images of traditional jammingsignals (including radio frequency (RF) noise, frequency-modulation (FM) noise, amplitude-modulation (AM) noise,constant range gate pull off (RGPO), velocity gate pull off(VGPO), convolutional modulation (CM), intermittent sam-pling (IS)) and the time and frequency domain images of noveljamming signals (including smeared spectrum (SMSP), chop-ping and interleaving (CI), smart noise jamming (SNJ), rangedeception jamming signal - amplitude modulation noise (RD-AM), and range deception jamming - frequency modulationnoise (RD-FM)) are shown in Fig. 19 and Fig. 20 respectively. (cid:11)(cid:68)(cid:12) (cid:11)(cid:69)(cid:12) (cid:11)(cid:70)(cid:12)(cid:11)(cid:71)(cid:12) (cid:11)(cid:72)(cid:12) (cid:11)(cid:73)(cid:12)(cid:11) (cid:12)(cid:11) (cid:12)(cid:11)(cid:71)(cid:12) (cid:11) (cid:12)(cid:11) (cid:12) (cid:11) (cid:12)(cid:11)(cid:73)(cid:12) (cid:11)(cid:74)(cid:12)
Fig. 19. The jamming signals(2D time-frequency images), (a)RF, (b)FM,(c)AM, (d)RGPO, (e)VGPO, (f)CM, (g)IS.
A. Jamming or Interference Classification and Recognition
Jamming or interference recognition is very important inradar target detection, tracking, recognition, and anti-jamming
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 30 or interference suppression tasks. In [559], the authors em-ployed SVM classifier to classify six different types of radar-to-radar interference waveforms, including time-frequency do-main signal and range-doppler profiles of different typesof interference. As the black-and-white problem of jammerclassification in Global Navigation Satellite System (GNSS),SVM and CNN were used to classify six types jammed signals,which achieved an accuracy rate of 94.90% and 91.36%,respectively in [560]. A method of dense false targets jammingrecognition was proposed in [561], which was based on a Ga-bor time-frequency atomic decomposition and SVM classifier.SVMs were also used to recognize the satellite interference[563] and radio ground-to-air interference signals [564]. Arobust neural network classifier was employed in [562] forclassification of frequency-modulated wideband radar jammersignals. In [565], the neural networks were also developedto recognize compound jamming signals based on featuresextracted in time domain, frequency domain and fractal di-mensions. These signals included additive, multiplicative andconvolution signals of typical blanket jamming and deceptionjamming.DL algorithms were also exploited to apply in jammingsignals classification and recognition. In [566], the authorsproposed an automatic jamming signal classification methodbased on CNN model, including audio jamming, narrowbandjamming, pulse jamming, sweep jamming and spread spectrumjamming. CNN model, based on time-frequency image as theinputs, was developed in [567] to classify 9 typical jammingswith an accuracy rate of 98.667% under the jammer-to-noiseratio (JNR) of 0dB-8dB. A LeNet-5 model based on spectrumwaterfall was proposed to recognize the jamming patterns in[568]. Similarly, a fine tuning LeNet, with 1D sequences (sizeof 1*896) as inputs, also employed for 7 kinds of jammingsidentification in [569], which achieved an accuracy rate of98%. In [570], a DL architecture was proposed to identifythe jamming factors of electronic information system. Therecognition method of four active jamming signal, based onCNN and STFT images as inputs, was proposed in [571],which achieved an accuracy rate of 99.86%, including blan-ket jamming, multiple false target jamming, narrow pulsejamming, and pure signal. The shadow features based onCNN algorithm were proposed for SAR deception jammingrecognition in [572]. As a multi-user automatic modulationclassification task, compound jamming signals recognitionbased on multi-label CNN model was proposed in [573]. Inaddition, a jamming prediction method based on DNN andLSTM algorithm was proposed in [574]. The jamming featuresextracted from PWDs list by DNN and were as the inputs ofLSTM for jamming prediction. The AE network consisted ofseveral layers of RNNs was proposed to detect interferencesignals based on time-frequency images in [575].
B. Anti-jamming and Interference Mitigation Strategies
As a strategy-making process, RL algorithms are usuallyadopted to make anti-jamming and interference strategies fordesigning intelligent algorithms. A DQN-based Q-learningalgorithm was employed in [576] to learn the jammer’sstrategies to design optimal frequency hopping strategies as (cid:11)(cid:68)(cid:12) (cid:11)(cid:69)(cid:12) (cid:11)(cid:70)(cid:12)(cid:11)(cid:71)(cid:12)
Fig. 20. The novel jamming signals, time domain (left) and frequency(right), (a)SMSP, (b)CI, (c)SNJ, (d)RD-AM. the anti-jamming strategy of cognitive radar. Spatial anti-jamming scheme for Internet of satellites based on deepRL and Stackelberg game was proposed in [577], whichregarded routing anti-jamming problem as a hierarchical anti-jamming Stackelberg game. The available routing subsets forfast anti-jamming decision-making were determined by deepRL algorithm to meet high dynamics caused by the unknowninterrupts and the unknown congestion.As for interference mitigation, a decentralized spectrumallocation strategy was developed in [578], which was basedon RL and LSTM model to avoid mutual interference amongautomotive radars. LSTM was used to aggregate the radar’sobservations for obtaining more information contributed to RLalgorithm. Similarly, a GRU-based RNN algorithm was usedfor interference mitigation of automotive radar in [579].VI. O
THER
ML-
BASED
RSP-
RELATED R ESEARCH
In addition to the applications detailed in sections III-V,there are some other that are worth , such as radar waveformoptimization design by RL [580], [581], radar spectrum alloca-tion [578], [582], [583], CEW [584], cognitive radar detection[585], antenna array selection via DL [586], and moving targetindication (MTI) using CNN [587].Compared to DL, RL performs well when used for cognitivedecision-making. Therefore, RL is suitable for strategy-makingbased RSP and radar system design, such as waveform design,
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 31 anti-jamming, CEW. An intelligent waveform optimizationdesign based on RL method was developed in [580] for multi-target detection of MIMO radar. The sum of target detectionprobability in range and azimuth cells was as the reward ofthe learning agent in each learning iteration, to estimate thetarget detection range and azimuth information. The optimizedweighted vector matrix of the transmitted waveform was asthe action space of the learning agent. This novel method canimprove the performance in detection probability, compared toall-direction waveform design methods. In addition, an end-to-end learning method for joint design the waveform detectorwas proposed in [581]. Both transmitter and receiver were im-plemented as feedforward neural networks, while the detectorand the transmitted waveform were trained alternately. Thisalgorithm achieved better robustness in clutter and colorednoise scenario than traditional methods.In [584], the authors applied deep RL in CEW for targetsearching, which built a 3D simulation CEW environment toaddress the spatial sparsity, continuous action, and partiallyobservable environment existing in CEW. A method of ML-based adaptive optimal target detection threshold estimationin non-Gaussian clutter environment was proposed in [585],which was effective even when the clutter distribution isunknown. A DL method was used for phased array antennaselection to better estimate direction-of-arrival (DOA) in [586],which constructed a CNN as a multi-class classification frame-work. The network determined a new antenna subarray foreach radar echo data in a cognitive operation pattern. ACNN-MTI structure was developed in [587] to overcome theconstrains of STAP-MTI, which performed feature extractionand classification directly from airborne radar echo. CNN-MTIhas proven more robust compared to traditional STAP-CNNand POLY methods.VII. D
ISCUSSION AND F UTURE P ROMISING T RENDS
Most of the current researches effort has concentrated on theapplications of ML to classification and recognition problems.Nevertheless, ML can be further exploited for different RSPapplications, such as target detection and tracking. Moreover,unified architectures for detection, tracking and recognitionmay be conceived that exploit ML as a common framework.This chapter will firstly put forward the future research di-rections, i.e., possible promising research topics. Then, thedetail research contents related to above research topics willbe profoundly discussed.
A. Research Topics
1) End-to-End Unified Intelligent Detection Architec-ture .In [589], the authors certified that the outputs of neuralnetwork with MSE or cross entropy as loss function sat-isfied the Neyman-Pearson detection criterion. Therefore, itis promising to exploit an intelligent end-to-end architectureby taking full use of the general non-linear fitting ability ofDNNs for radar target detection. The functions of this schemeinclude pulse compression, coherent accumulation, and CFARin an unified end-to-end learning manner. The challengeable research problems include the intelligent CFAR, environmentidentification (such as noise and clutter background automaticclassification [588]) techniques. For example, because of theextensive distributed mapping ability of RNN with attentionmechanism, problems such as target sidelobe masking, multi-target interference, and target model mismatch may be solvedusing RNN-related architecture.
2) Target Detection and Tracking-Unified IntelligentProcessing .Building an effective closed loop network of unified targetdetection and tracking can improve the performance of stabletarget tracking with clutter in the background. It is important tostudy on the performance evaluation metrics and parameter op-timization techniques of target detection and tracking. For ex-ample, it is possible to optimally adjust the detection thresholdvia prior knowledge-based online learning techniques, which isbased on the feedback from target tracking information (suchas motion trends, covariance estimation) to target detectionunits. This operation maybe contribute to track the target flighttrajectory and improve the detection probability of subsequentpoint trajectory for confirmed targets.
3) End-to-End Framework of Unified Target Detection,Tracking and Recognition .Based on previous two research, it is promising to studyan end-to-end architecture to achieve unified intelligent pro-cessing for clutter suppression, radar target detection, tracking,and recognition by ANNs-based multi-task learning. Becauseof powerful non-linear fitting ability, ANNs have high per-formance in classification and recognition tasks. Accordingto the targets (valuable targets, clutter or noise background)recognition information, radar can program the optimal track-ing route based on the ANNs-based prediction of target flighttrajectory. In addition, target detection is a special type oftarget recognition, therefore, it can assist target detection task.However, it is extremely changeable about how to effectivelybuild integrated signal processing mechanism and detectionframework, which can promote each other, uniformly makedecision-making.
B. Promising Research Contents
1) The Solutions of the Limitation of Dataset .Classification and recognition of radar targets suffers ofthe typical problem of a small amount of labeled samples.To improve the performance with limited data samples, itis necessary to augment the limited data or design effectivelearning algorithms with limited data. In addition, in orderto reduce the cost of obtaining real data, simulation dataset,closely to simulate real complex electromagnetic environment,is also needed to train DL model by transfer learning pattern.
Data augmentation
The existing data augmentation meth-ods mainly focus on the manipulation of original data samples,e.g., manual extraction of sub-images, add noise, filtering, andflipping [253]. In addition, a method of generating new datasamples with GANs was also used in [251], [252]. Practicaloperational conditions, however, are usually neglected whenapplying these methods to some extent, which make the newdata retain the same characteristics. Environmental conditionsare a significant component in the radar echo signal, such as
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 32 different scattering centers with different illuminating direc-tions, which produce different amplitude and phases in theecho signal. Therefore, data augmentation techniques withconsideration of radar practical operational conditions (such asSOCs and EOCs) are required. For example, the combinationof electromagnetic scattering computation with dynamic orstatic measurements may be used to improve the accuracyand robustness of target recognition algorithms. Moreover,exploiting the evaluation metrics of generated data equality toefficiently assist the data generations. In this way, the learningmodel can learn more discriminative features of unknowntargets and improve performances in terms of accuracy andgeneralization.
Few/zero shot learning
This research direction mainlyexploits how to effectively extract discriminative features outof small training samples, to improve accuracy and general-ization performances. At present, some achievements in thisdirection have been obtained, such as the design efficientlearning model. For example, a feature fusion frameworkwas presented in [255] based on the Gabor features andinformation of raw SAR images, to fuse the feature vectorsextracted from different layers of the proposed neural network.A TL method was employed in [256] to transfer knowledgelearned from sufficient unlabeled SAR scene images intolabeled SAR target data. A-ConvNets was proposed in [259]to significantly reduce the number of free parameters andthe degree of overfitting with small datasets. Group squeezeexcitation sparsely connected CNN was developed in [346]to perform dynamic channel-wise feature recalibration withless model parameters. To balance the feature extraction andanti-overfitting, a CAE-HL-CNN model was proposed in [412]to perform effective learning for the classification of MSTARtargets with small training samples.However, this research direction is just at the beginning, andneeds to have prolonged insight into it. Based on few/zero-shotlearning methods from the DL domain, some works [590]–[593] has been done to design effective learning algorithmsto address learning issues with small data samples. Few-shotlearning can rapidly generalize to new tasks of limited super-vised experience by turning to prior knowledge, which mimicshuman’s ability to acquire knowledge from few examplesthrough generalization and analogy [590]. Zero-shot learningaims to precisely recognize unseen categories through a sharedvisual-semantic function, which is built on the seen categoriesand expected to well adapt to unseen categories [593].
2) The Design of Lightweight Algorithms .Since the requirements of real-time signal processing andhigh sampling rate in RSP domain are quite demanding, alarge volume of parameters of DL model is still a severechallenge for real-time optimal training, which results in highstorage and computation complexity. Lightweight DL modelshave been proposed, such as MobileNets (v1-v3) [87]–[89]),ShuffleNets (v1-v2) [90], [91]), to be embedded in mobilephones or other portable device. Nevertheless, these modelsare do not fully meet the requirements of RSP, such as lowmemory resource and strict latency requirements. Therefore,the design of lightweight models or efficient DNN models[39] is necessary for DL model to be efficiently applied to RSP domain. Research on novel lightweight architecture design anddeep model compression and accelerating methods, specializedfor RSP, is mandatory for enabling this technology.
Neural architecture search (NAS)
Currently employedarchitectures in DL have mostly been developed manually byhuman experts, which is a time-consuming and error-proneprocess.
NAS provides a promising solution to alleviate thisissue [594], being an automatical architecture design system.NAS includes three steps: search space , search strategy , and performance estimation . The purpose of NAS is typicallyto find the best architectures from a search space that canhighly achieve predictive performance on unknown data. Theapplication of NAS to identify optimal architectures for RSPis an interesting future trend. Deep model compression and accelerating methods
DNNshave achieved great success in many CV tasks. Computationcosts and storage intensity, however, are the main challenges ofexisting DNN models that hinder their deployment in deviceswith low memory resources or in applications with strict la-tency requirements. Deep model compression and acceleratingmethods have been developed to address these challenges inrecent years, including parameter pruning and sharing, low-rank factorization, transferred/compact convolutional filters,and knowledge distillation [41], [595]. These methods are stillin the early stages, as most of techniques only aim at CNNmodels and classification tasks. It is critical to extensivelydevelop these compression and accelerating techniques in RSPdomain.
3) Explainable Machine Learning Algorithms .ML has achieved great success in many domains. Black-boxproperty of the DNN model [596], however, demonstrates asevere challenge in practical applications of ML algorithms,such as medical image precessing domain and bank investmentdecision making. For example, a special doctor needs toclearly know the model how to make decisions in a explain-able manner. When we do some ML-based works, especiallyDNNs, some questions naturally emerge in our mind. Forexample, how does the network model work? What does theinner structure of the model do about the inputs? Why doesthe model have the ability to classify, detect, liked humanbrain does? These questions are triggered by the black boxproperty of non-interpretable ML models. Therefore, to widelyapply ML in practice, explainable artificial intelligence (XAI)is a key factor [597], [598]. Especially, XAI techniques areextremely important in RSP, such as the classification andrecognition of valuable targets and recognition of jammingtypes in situation awareness domain. A military commanderneeds to clearly understand the process of decision-makingof ML models to believe the model, so as to deploy highlyeffective decision strategies, e.g., anti-jamming countermea-sures, weapon deployment strategies in electronic warfare.The present published literatures about XAI can be roughlycategorized into four classes:i) post-hoc explanations, such as local/global proxy mod-els assisting explanation [599]–[601], visualization of latentpresentation [602]–[604], analysis of attributes for prediction[605], [606];
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 33 ii) transparent/interpretable model design, such as embed-ding transparent/interpretable model in DNNs [607]–[609],regularization-based design [609]–[611], disentangled repre-sentation learning [612], [613], attention mechanism baseddesign [614], [615];iii) interdisciplinary knowledge embedded, such as informa-tion theory [616], [617], physics theory [618], [619], humanbrain cognitive science [620]–[622];iv) combining symbolism and connectionism [623], [624].In this section, we will take a brief discussion for combiningsymbolism and connections as an extremely promising methodfor XAI. The present ML algorithms are just used in aninformation sensing (i.e., pattern recognition) domain to largeextent, which do not have the abilities of casual reasoning[625] and interpretability. Therefore, it has many challenges inpractical applications, such as the requirement of large trainingdata, overfitting, robustness, adversarial attacks [40]. The deepmodel has an excellent ability to learn features with largedataset. These features, however, are usually high dimensionaland redundant.Good representation of data should be that these features,extracted from the learning model, are low-dimensional, ab-stract, and discrete [626], i.e., concept features [622], whichare similar to the characteristics of the symbolism learningrepresentation method. Symbolism learning [627] has theability, through logical reasoning, to produce semantic fea-tures, which can be logically understood by human beings.Therefore, it is possible to combine symbolism learning (withthe ability of logical reasoning) with connectionism learning(i.e., deep learning with ability of powerful features extraction)to achieve human-level concept learning [622]. Yoshua Bengiohas recently proposed consciousness prior as a suitable toolto bridge the gap between the symbolism and connectionism[628], which can combine attention mechanism to extractconsciousness features from semantic features of RNNs withconsciousness prior.
4) Cognitive Waveform Design .As the main situation awareness sensing system in EW,a radar system is primarily responsible for surveillance andtracking of the EW environment, including target detec-tion and recognition, jamming/interference countermeasures,and counter-countermeasures, which are consistent with themissions of EW, i.e., electronic support measure (ESM),electronic countermeasure (ECM), and electronic counter-countermeasure (ECCM) systems [629]. With the developmentof the cognitive electronic warfare (CEW) in recent years[584], many challenges have emerged that affect radar systems.As a possible solution, cognitive radar (CR) has been proposedin [630], which is an interdiscipline research domain of neuro-science, cognitive science, brain science and RSP [631]. Threebasic ingredients of CR are i) intelligent signal processing; ii)feedback from receiver to transmitter; and iii) preservation ofinformation content of radar returns [630]. The basic conceptsof CR mainly focus on knowledge-based adaptive radar [632].With the rapid development of ML, especially DL and RL,CR should have novel promising research contents based onadvanced ML algorithms in the future.Radar waveform design is one of the significant tasks in the design of radar system. Traditional radar usually transmits onlyone or few types of waveforms to optimise target detection.As a key task of CR, cognitive waveform optimization designhas attracted a lot of attention [633]. CR makes full use ofthe knowledge of the external environment and targets, todesign optimal waveforms to optimise the tasks of target de-tection, anti-jamming/interfenrence, at the conditions of radarconstraints, objective optimization principles, and advancedoptimization theory. The optimization problem of waveformdesign, however, is a non-convex, high dimension, and multi-constraint optimization problem, whose global optimal solu-tion is usually difficult to find at low computational costs. ML-based optimization methods may indicate alternative directionsto address this challenge. Moreover, the optimization processis an iterative search procedure to find the optimal solution,which can be regarded as a problem of sequence decision-making. RNN and RL are good at sequential data processingand optimal strategy-making, respectively. Therefore, it ispossible to combine optimization theory with ML to improvethe performance in radar waveform optimization design. Someinitial works about this theme have emerged, such as branch-and-bound algorithm of mixed-integer linear programmingwith ML technologies in [634]–[636], ML for combinatorialoptimization [637], [638], RL for solving the vehicle routingproblem [639], and pointer networks with one-layer RNN[640].
5) Intelligent Anti-jamming .The efficient anti-jamming techniques are concerned withtoward challenges in an increasingly complex electromagneticenvironment. It is difficult for traditional anti-jamming tech-niques to face current requirements of modern radar systemsequipments. The vision of intelligent anti-jamming methods isincreasingly intensive with the rapid development of artificialintelligence. In recent years, a new research wave has advancedin this field based on ML algorithms, such as RL-based anti-jamming or interference [576], [577]. This new direction needsto be deeply exploited to address the existing challenges,including jamming recognition, anti-jamming strategy, and thedefinition of performance metrics.
Jamming recognition
This aspect has been deeply discussedin the first three parts, which is similar to radar targetrecognition tasks. Multi-task, multi-view, multi-scale learningtechniques, and attention mechanism learning method shouldbe considered.
Anti-jamming strategy
The selection of efficient anti-jamming measurements is a decision-making process. DeepRL seems to be a promising research lead when training anagent to automatically select adaptive anti-jamming measureswith the assistance of knowledge of external environment andtargets.
Performance Evaluation metrics
Although some RL-basedachievements in terms of intelligent anti-jamming have beenreached, there is little research done in terms of performanceevaluation metrics. This direction is vital to evaluate theperformances of anti-jamming techniques, which also canassist to select optimal anti-jamming measures.
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 34
VIII. C
ONCLUSION
There is a strong evidence of the extensive developmentof ML-based RSP algorithms that have found applicationin several radar-related fields. Some areas seem to be moretargeted than others due to the direct application of ML-based techniques and because of the strong interest of manyresearchers, which is likely driven by strong interests fromstakeholders. Particularly radar image processing and relativeclassification is one area where ML-based algorithms mayprove a valid solution to current challenges. In this paper, wehave provided a structured and amply commented literaturesurvey, followed by indications about future leads, which maybe used by many researchers and practitioners to inspire themand help them progressing with their work in this field.A
PPENDIX AL IST OF ACRONYMS
ACF autocorrelation functionAWGN additive white Gaussian noiseASCs attributed scattering centersAdaboost adaptive boostingANN artificial neural networkAEs autoencodersATR automatic target recognitionARPNs advanced region proposal networkBCRN bidirectional convolution-recurrent networkBM3D blocking matching 3DCNNs convolutional neural networksCEW cognitive electronic warfareCDAE convolutional denoising autoencoderCV computer visionCOCO common objects in contextCWNN convolutional wavelet neural networkCTFD Cohen’s time-frequency distributionCVDFP cadence velocity diagram frequency profileCR cognitive radarCV-CNN complex-value CNNCFAR constant false alarm ratesCPON class probability output networkCWTFD Choi-Williams time-frequency distributionConvLSTM convolutional LSTMCS compressive sensingDT decision treeDis-iRBM discriminative infinite RBMDeAE denoising autoencoderDRL deep reinforcement learningDBNs deep belief networksDLFM dual linear frequency modulationDNNs deep neural networksDCGANs Deep convolutional GANsDQN deep Q networkDCC-CNNs despeckling and classification coupled CNNsDCFM-CNN dual channel feature mapping CNNDARPA Defense Advanced Research Projects AgencyDSC depthwise separable convolutionEW electronic warfareEL Ensemble Learning ESM electronic support measureECM electronic countermeasureECCM electronic counter-countermeasureECOs extended operating conditionsEQFM even quadratic frequency modulationENN Elman neural networkFCN fully convolutional networkFCBF fast correlation-based filterFPNs feature pyramid networksFRN-MSF feature recalibration network with multi-scalespatial featuresGNN graphical neural networkGIoU generalized intersection over unionGF-3 Gaofen-3GRU gated recurrent unitGBDT gradient boosting decision treeGANs generative adversarial networksGPUs graphical processing unitsGCN global convolutional networkHAPs high-altitude platformsHMMs Hidden Markov ModelsHR high-resolutionHRRP high resolution range profileHL hinge lossIDCNN image despeckling convolutional neural networkInSAR interferometric SARIDCNN image despeckling convolutional neural networkILSVRC ImageNet large-scale visual recognition challengeICS iterative censoring schemeISAR inverse synthetic aperture radarIoU intersection of unionJSDC joint supervised dictionary and classifierJNR jammer-to-noise ratioK-NN K-nearest neighborKSR kernel sparse representationLSTM long short-term memoryLIME local interpretable model-agnostic explanationsLSGANs least squares generative adversarial networksLFM linear frequency modulationLFO-Net lightweight feature optimizing networkLPI low probability interceptLFMCW linear frequency modulation continuous waveLOS line of sightML machine learningMS-CNN multi-stream CNNMDP Markov decision processMSTAR moving and stationary target acquisition and recog-nitionMP mono-pulseMLP multi-layer perceptronMTI moving target indicationMRF Markov random fieldMLFM multiple linear frequency modulationMIMO multi-input and multi-outputMVL multi-view learningMAP mean average precisionMTL multi-task learning
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 35
MSCDAE modified stacked convolutional denoising auto-encoderMPDPL multilayer projective dictionary pair learningMRDED multi-resolution dense encoder and decoderNLP natural language processingNN neural networkNAS neural architecture searchNMS non-maximum suppressionOMMPS orthogonal maximum margin projection subspaceOWN optimized Wishart networPRI pulse repetition intervalPRI Pulse repetition intervalPSR probability of successful recognitionPTCNN probability transition convolutional neural networkPWDs pulse description wordsPPCA probabilistic principal component analysisPCA principal component analysisPSDNN patch-sorted deep neural networkPGBN poisson gamma belief networkPSF point spread functionRSP radar signal processingRFMLS radio frequency machine learning systemRRSCR radar radiation sources classification and recogni-tionRVM relevant vector machineRL reinforcement learningRF random forestR-CNN regional convolutional neural networkRFI radio frequency identificationRMA range migration algorithmRNNs recurrent neural networksRS remote sensingRESISC remote sensing image scene classificationRBF radial basis functionREC radar emitter classificationRBM restricted Boltzmann machineRVFL random vector functional linkRAD range doppler algorithmSSP speech signal processingSFM sinusoidal frequency modulationSAR synthetic aperture radarSEI specific emitter identificationSCR signal-to-clutter ratioSTAP-MTI spatial time adaptive processing and motiontarget indicationSSD shot multiBox detectorSTD ship targets detectionSNR signal noise ratioSE-Net sequeeze-and-excitation networkSVMs support vector machinesSAR-DRN SAR dilated residual networkSCCs sparse coding-based classifiersSOCs standard operating conditionsSMO sequence minimization optimizationSAE sparse autoencoderSVD single vector decompositionSFP spectrogram frequency profileSTFT short time fourier transformation SPP spatial pyramid poolingSRDNN superpixel restrained DNNSBL sparse Bayesian learningTL transfer learningTARAN Target-aware recurrent attentional networkTPLBP three patch local binary patternt-SNE t-distributed stochastic neighbor embeddingTFD time-frequency distributionTPOT tree-based pipeline optimization toolTOA time of arrivalTLM texture level mapTPUs tensor processing unitsU-CNN unidimensional convolutional neural networkUAV unmanned aerial vehicleVAE variational autoencoderWGAN Wasserstein GANWGAN-GP Wasserstein GAN with a gradient penaltyWKCNN weighted kernel CNNWKM weighted kernel moduleXGBoost extreme gradient boosting decision treeXAI explainable artificial intelligenceYOLO You Only Look OnceR
EFERENCES[1] George W. Stimson, Introduction to airborne radar, second edition,SciTech Publishing, Inc.ISBN: 1-891121-01-4, 1998.[2] Bassem R. Mahafza, Radar systems analysis and design using MAT-LAB, Chapman and Hall/CRC Publishing, 2000.[3] Mark A.Richards, Fundamentals of radar signal processing, secondedition, McGraw-Hill Publishing, ISBN 13: 9780071798327, 2014.[4] Sergei A. Vakin, Lev N. Shustov, Robert H. Dunwell, Fundamentals ofelectronic warfare, Artech House Publishing, Inc.ISBN: 1-58053-052-4, 2001.[5] D. Yu, L. Deng, “Deep learning and its applications to signaland information processing [Exploratory DSP],”
IEEE Signal Pro-cessing Magazine , vol. 28, no. 1, pp. 145-154, Jan. 2011. Doi:10.1109/MSP.2010.939038.[6] Christopher M. Bishop, Pattern recognition and machine learning,Springer India Publishing, ISBN: 9788132209065, 2013.[7] I.Goodfellow, Y. Bengio, A. Courville, Deep learning, MIT Press,ISBN: 9780262035613, 2016.[8] LeCun, Y., Bengio, Y., Hinton, G., “Deep learning,” Nature, 521 (7553),pp. 436 − IEEE Signal Processing Magazine , vol.34, no. 6, pp. 26-38, 2017.
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 36 [17] W. G. Hatcher and W. Yu, “A survey of deep learning: platforms,applications and emerging research trends,”
IEEE Access , vol.6, pp.24411-24432, 2018.[18] Tan C, Sun F, Kong T, et al., “A survey on deep transfer learning,”in
Proc. International Conference on Artificial Neural Networks , 2018,PP. 270-279.[19] L. Weibo, W. Zidong, L. Xiaohui, et al., “A survey of deep neuralnetwork architectures and their applications,”
Neurocomputing , vol.234, pp. 11-26, 2017.[20] Z. Pan, W. Yu, X. Yi, et al., “Recent progress on generative adversarialnetworks (GANs): A survey,”
IEEE Access , pp. 36322-36333, 2019.[21] G. J. Litjens, T. Kooi, B.E. Bejnordi, et al., “A survey on deep learningin medical image analysis,”
Medical Image Analysis , pp. 60-88, 2017.[22] R. J. G. van Sloun, R. Cohen, Y. C. Eldar, “Deep learning in ultrasoundimaging,”
Proceedings of the IEEE , vol. 108, no. 1, pp. 11-29, Jan.2020. Doi: 10.1109/JPROC.2019.2932116.[23] A. Kamilaris, F.X. Prenafetaboldu, “Deep learning in agriculture: Asurvey,”
Computers and Electronics in Agriculture , pp. 70-90, 2018.[24] L. Zhang, S. Wang, B. Liu, et al., “Deep learning for sentimentanalysis: A survey,”
Wiley Interdisciplinary Reviews-Data Mining andKnowledge Discovery , vol. 8, no. 4, 2018.[25] Sabour S, Frosst N, Hinton G E, et al, “ Dynamic routing betweencapsules,” arXiv: Computer Vision and Pattern Recognition , 2017.[26] A. Harry. S. Pierson, Michael, Gashler, “Deep learning in robotics: Areview of recent research,”
ArXiv:
IEEE Geoscience andRemote Sensing Magazine , vol. 5, no. 4, pp. 8-36, 2017.[28] L. Zhang, L.Zhang, B. Du, et al., “Deep learning for remote sensingdata: A technical tutorial on the state of the art,”
IEEE Geoscience andRemote Sensing Magazine , vol. 4, no. 2, pp. 22-40, 2016. .[29] J. E. Ball, D. T. Anderson, C. S. Chan, et al. “Comprehensive surveyof deep learning in remote sensing: Theories, tools, and challenges forthe community,”
Journal of Applied Remote Sensing , vol. 11. no. 4,2017.[30] A. E. Maxwell, T. A. Warner, F. Fang, et al., “Implementation ofmachine-learning classification in remote sensing: An applied review,”
International Journal of Remote Sensing , vol. 39, no. 9, pp. 2784-2817,2018.[31] L. Ma, Y. Liu, X. Zhang, et al., “Deep learning in remote sensing appli-cations: A meta-analysis and review,”
Isprs Journal of Photogrammetryand Remote Sensing , pp. 166-177, 2019.[32] G. Cheng, J. Han, X. Lu, et al., “Remote sensing image sceneclassification: Benchmark and state of the art,” in
Proc. ComputerVision and Pattern Recognition, (CVPR) , 2017, pp. 1865-1883.[33] G. Dong, G. Liao, H. Liu, et al., “A review of the autoencoderand its variants: A comparative perspective from target recognitionin synthetic-aperture radar images,”
IEEE Geoscience and RemoteSensing Magazine , vol. 6, no. 3, pp. 44-68, 2018.[34] K. El-Darymli, E. W. Gill, P. Mcguire, et al., “Automatic targetrecognition in synthetic aperture radar imagery: A state-of-the-artreview,”
IEEE Access , vol. 4, pp. 6014-6058, 2016.[35] U. Kanjir, H. Greidanus, K. Ostir, “Vessel detection and clas-sification from spaceborne optical images: A literature survey,”
Remote Sensing of Environment , vol. 207, pp. 1-26, 2018.https://doi.org/10.1016/j.rse.2017.12.033.[36] Y. Li, M. Yang, Z. Zhang, et al., “A survey of multi-view representationlearning,”
IEEE Transactions on Knowledge and Data Engineering ,vol. 31, no. 10, pp. 1863-1883, 2019.[37] S. Ruder, A. Ltd, Dublin, “An overview of multi-task learning in deepneural network,”
ArXiv : 170605089, 2017.[38] Z. Li, W. Yang, S. Peng, et al., “A survey of convolutional neuralnetworks: Analysis, applications, and prospects,”
ArXiv : 2004.02806v1,2020.[39] V. Sze, Y. Chen, T. Yang, J. S. Emer, “Efficient processing of deep neu-ral networks: A tutorial and survey,”
Proceedings of the IEEE , vol. 105,no. 12, pp. 2295-2329, Dec. 2017, doi: 10.1109/JPROC.2017.2761740.[40] D. J. Miller, Z. Xiang, G. Kesidis, “Adversarial learning targeting deepneural network classification: A comprehensive review of defensesagainst attacks,”
Proceedings of the IEEE , vol. 108, no. 3, pp. 402-433, March 2020. Doi: 10.1109/JPROC.2020.2970615.[41] B. L. Deng, G. Li, S. Han, L. Shi, Y. Xie, “Model compression andhardware acceleration for neural networks: A comprehensive survey,”
Proceedings of the IEEE , vol. 108, no. 4, pp. 485-532, April 2020.Doi: 10.1109/JPROC.2020.2976475. [42] Z. Zhao, P. Zheng, S. Xu, et al., “Object detection with deep learning: Areview,”
IEEE Transactions on Neural Networks and Learning Systems ,vol. 30, no. 11, pp. 3212-3232, Nov. 2019.[43] L. Liu, W. Ouyang, X. Wang, et al., Deep learning for generic objectdetection: A survey,”
International Journal of Computer Vision , vol.128, no. 2, pp. 261-318, 2020.[44] Burges C J, “A tutorial on support vector machines for patternrecognition,”
Data Mining and Knowledge Discovery , vol. 2, no. 2,pp. 121-167, 1998.[45] T. Zhang, “An introduction to support vector machines and otherkernel-based learning methods,”
Ai Magazine , vol. 22, no. 2, pp. 103-104, 2001.[46] J. Fleiss, B. Levin, M. Paik, “Logistic Regression,”10.1002/0471445428.ch11, 2004.[47] T. Long, T. Yingjie, Y. Chunyan, “Nonparallel support vector regressionmodel and its SMO-type solver,”
Neural Networks , vol. 105, pp. 431-446, 2018.[48] J. R. Quinlan, “Induction of decision trees,”
Machine Learning , vol. 1,no. 1, pp. 81-106, 1986.[49] J. R. Quinlan, “C4.5: Programs for machine learning,” Morgan Kauff-man, 1993.[50] W. Loh, “Classification and regression trees,”
Wiley InterdisciplinaryReviews-Data Mining and Knowledge Discovery , vol. 1, no. 1, pp. 14-23, 2011. https://doi.org/10.1002/widm.8.[51] T. G. Dietterich, “Ensemble methods in machine learning,”
MultipleClassifier Systems , pp. 1-15, 2000.[52] B. Leo, “Random Forests,”
Machine Language , 2001.https://doi.org/10.1023/A:1010933404324[53] Y. Sun, Z. Liu, S. Todorovic et al., “Adaptive boosting for SARautomatic target recognition,”
IEEE Transactions on Aerospace andElectronic Systems , vol. 43, no. 1, pp. 112-125, Jan. 2007.[54] C. Tianqi, G. Carlos, “XGBoost: A scalable tree boosting system,”
ArXiv :1603.02754v1, 2016.[55] J. Friedman. “Greedy function approximation: A gradient boostingmachine,”
Annals of Statistics , pp. 1189-1232, 2001.[56] F. M. Foody, M. K. Arora, “An evaluation of some factors afectingthe accuracy of classification by an artificial neural network,”
In-ternational Journal of Remote Sensing vol. 18, pp. 799-810, 1997.Doi:10.1080/014311697218764.[57] E. Leif Peterson, “K-nearest neighbor,”
Scholarpedia , vol. 4, no. 2,1883. doi:10.4249/scholarpedia.[58] G. E. Hinton, S. Osindero Y. Teh, “A fast learning algorithm for deepbelief nets,”
Neural Computation , vol. 18, no. 7, pp. 1527-1554, Jul.2006.[59] A. Krizhevsky, I. Sutskever, G. Hinton, 2012. “Imagenet classificationwith deep convolutional neural networks,”
Advances in Neural Infor-mation Processing Systems , pp. 1097-1105, 2012.[60] L. Deng, J. Li, J. Huang, et al., “Recent advances in deep learningfor speech research at Microsoft,” in
Proc. International conference onacoustics, speech, and signal processing , 2013, pp. 8604-8608.[61] C. Szegedy, A. Toshev, D. Erhan, et al., “Deep neural networks forobject detection,” in
Proc. Neural Information Processing Systems ,2013, pp. 2553-2561.[62] B, Zhou, A. Lapedriza, A. Khosla, et al., “Places: A 10 million imagedatabase for scene recognition,”
IEEE Transactions on Pattern Analysisand Machine Intelligence , vol. 40, no. 6, pp. 1452-1464, 2018.[63] A. Garciagarcia, S. Ortsescolano, S. Oprea, et al., “A review ondeep learning techniques applied to semantic segmentation,” in
Proc.Computer Vision and Pattern Recognition (CVPR) , 2017.[64] A. Mohamed, G. E. Dahl, G. E. Hinton, et al., “Acoustic modelingusing deep belief networks,”
IEEE Transactions on Audio, Speech, andLanguage Processing , vol. 20, no. 1, pp. 14-22, 2012.[65] A. V., S. Dieleman, B. Schrauwen, et al., “Deep content-based musicrecommendation,” in
Proc. Neural Information Processing Systems ,2013, pp. 2643-2651.[66] Q. Abbas, M. E. Ibrahim, M. A. Jaffar, et al., “Video scene analysis:An overview and challenges on deep learning algorithms,”
MultimediaTools and Applications , vol. 77, no. 16, pp. 20415-20453, 2018.[67] L. Wang, D. Sng, “Deep learning algorithms with applications to videoanalytics for a smart city: A survey,” in
Proc. Computer Vision andPattern Recognition , 2015.[68] Y. Xie, L. Le, Y. Zhou, et al., “Deep learning for natural languageprocessing,”
Handbook of Statistics , 2018.[69] M. M. Lopez, J. K. Kalita, “Deep Learning applied to NLP,”
Compu-tation and Language , 2017.[70] T. Young, D. Hazarika, S. Poria, et al., “Recent trends in deep learningbased natural language processing,”
Computation and Language , 2017.
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 37 [71] H. Lee, R. Grosse, R. Ranganath. et al., “Convolutional deep beliefnetworks for scalable unsupervised learning of hierarchical representa-tions,” in
Proc. International Conference on Machine Learning (ICML) ,2009, 609-616.[72] N. Srivastava, R. Salakhutdinov, “Multimodal learning with deepboltzmann machines,” in
Proc. Neural Information Processing Systems ,2012, pp. 2222-2230.[73] E. Geoffrey, Hinton, “Deep belief networks,”. Scholarpedia, vol. 4, no.5, pp. 5947, 2009.[74] M. Ranzato, C. Poultney, S. Chopra, et al., “Efficient learning of sparserepresentations with an energy-based model,” in Proc. Advances inNeural Information Processing Systems, Vancouver, Canada, 2007, pp.1137-1144.[75] D. Kingma, S. Mohamed, D. Rezende, et al., “Semisupervised learningwith deep generative models,” in
Proc. Advances in Neural InformationProcessing Systems , Montreal, Canada, 2014, pp. 3581-3589.[76] D. Rezende, S. Mohamed, and D. Wierstra, “Stochastic backpropaga-tion and approximate inference in deep generative models,” in
Proc.ACM 31st Int. Conf. Machine Learning , Jun. 2014, pp. 1278-1286.[77] S. Rifai, P. Vincent, X. Muller, X. Glorot, and Y. Bengio, “Contractiveauto-encoders: Explicit invariance during feature extraction,” in
Proc.ACM 28th Int. Conf. Machine Learning , Bellevue, Washington, 2011,pp. 833-840.[78] J. Masci, U. Meier, D. Cirean, and J. Schmidhuber, “Stacked convo-lutional auto-encoders for hierarchical feature extraction,” in
Proc. Int.Conf. Artificial Neural Networks , 2011, pp. 52-59.[79] B. L. Lecun Y , Bengio Y , et al., “Gradient-based learning applied todocument recognition,”
IEEE Proceedings , vol. 86, no. 11, pp. 2278-2324, 1998.[80] A. Krizhevsky, I. Sutskever, and G. Hinton, “ImageNet classificationwith deep convolutional neural networks,”
Advances in neural infor-mation processing systems , vol. 25, no. 2, 2012.[81] K. Simonyan, and A. Zisserman, “Very deep convolutional networksfor large-scale image recognition,”
Computer Science , 2014.[82] Szegedy C, Liu W, Jia Y, et al, “Going deeper with convolutions,” in
Proc. Computer Vision and Pattern Recognition , 2015, pp. 1-9.[83] S. Ioffe, and C. Szegedy, “Batch normalization: Accelerating deep net-work training by reducing internal covariate shift,”
ArXiv : 1502.03167,2015.[84] C. Szegedy, V. Vanhoucke, Ioffe S, et al., “Rethinking the inceptionarchitecture for computer vision,” in
Proc. Computer Vision and PatternRecognition , 2016, pp. 2818-2826.[85] C. Szegedy, S. Ioffe, V. Vanhoucke, et al., “Inception-v4, inception-resNet and the impact of residual connections on learning,” in
Proc.National Conference on Artificial Intelligence , 2016, pp. 4278-4284.[86] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for imagerecognition,”
ArXiv : 1512.03385, 2015.[87] A. G. Howard, M. Zhu, B. Chen, et al, “Mobilenets: Efficient con-volutional neural networks for mobile vision applications,”
ArXiv :1704.04861, 2017.[88] M. Sandler, A. Howard, M. Zhu, et al, “Mobilenetv2: Inverted resid-uals and linear bottlenecks,” in
Proc. Computer Vision and PatternRecognition , 2018, pp. 4510-4520.[89] M. S. Andrew Howard, Grace Chu, Liang-Chieh Chen, et al., “Search-ing for MobileNetV3,”
ArXiv : 1905.02244, 2019.[90] Zhang X, Zhou X, Lin M, et al. “ShuffleNet: An extremely efficientconvolutional neural network for mobile devices,” in
Proc. ComputerVision and Pattern Recognition , 2018, pp. 6848-6856.[91] N. Ma, X. Zhang, H. Zheng, et al. “ShuffleNet V2: Practical guidelinesfor efficient CNN architecture design,” in
Proc. European Conferenceon Computer Vision , 2018, pp. 122-138.[92] T. Mingxing, V. L. Quoc,“ EfficientNet: Rethinking model scaling forconvolutional neural networks,”
ArXiv : 1905.11946, 2019.[93] K. Han, Y. Wang, Q. Tian, et al., “GhostNet: More features from cheapoperations,”
ArXiv : 1911.11907, 2019.[94] S. Hochreiter, J. Schmidhuber, “Long short-term memory,”
NeuralComputation , vol. 9, no. 8, pp. 1735-1780, 1997.[95] M. Ravanelli, P. Brakel, M. Omologo, et al., “Light gated recurrentunits for speech recognition,”
Audio and Speech Processing , vol. 2,no. 2, pp. 92-102, 2018.[96] I. Goodfellow, J. Pougetabadie, M. Mirza, et al., “Generative adversar-ial nets,” in
Proc. Neural Information Processing Systems , 2014, pp.2672-2680.[97] A. Creswell, T. White, V. Dumoulin, et al., “Generative adversarialnetworks: An overview,”
IEEE Signal Processing Magazine , vol. 35,no. 1, pp. 53-65, 2018. [98] Z, Wang, Q, She, T. E. Ward, et al., “Generative adversarial networks:A survey and taxonomy,”
ArXiv : Learning, 2019.[99] A. Radford, L. Metz, S. Chintala, et al., “Unsupervised representationlearning with deep convolutional generative adversarial networks,”
ArXiv : Learning, 2015.[100] M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein gan,”
ArXiv :1701.07875, 2017.[101] I. Gulrajani, F. Ahmed, M. Arjovsky, et al., “Improved training ofwasserstein GANS,” in
Proc. Advances Neural Information ProcessingSystems Conf. , 2017.[102] M. Mirza, S. Osindero, “Conditional generative adversarial nets,”
ArXiv : Learning, 2014.[103] J. Y. Zhu, T. Park, P. Isola, et al., “Unpaired image-to-image translationusing cycle-consistent adversarial networks,” in
Proc. InternationalConference on Computer Vision (ICCV) , 2017.[104] L. Yongyi, T. Yu-Wing, T. Chi-Keung, “Attribute-guided face genera-tion using conditional cycleGAN,
ArXiv : 1705.09966, 2018.[105] X. Chen, Y. Duan, R. Houthooft, et al., “InfoGAN: Interpretable rep-resentation learning by information maximizing generative adversarialnets,”
ArXiv : Learning, 2016.[106] A. E. Spezio, “Electronic warfare system,”
IEEE Trans. Microw. TheoryTech. , vol. 50, no. 3, pp. 633-644, 2002.[107] J. Zak, P. Zavorka, “The identification of radar signals by histogramdescription leads to cluster analysis processing,” in
Proc. 15th Int.Radar Symp. (IRS) , Dresden, Germany, 2014, pp. 1-4.[108] J. Matuszewski and L. Paradowski, “The knowledge based approachfor emitter identification,” in
Proc. 12th International Conference onMicrowaves and Radar , Krakow, Poland, 1998, pp. 810-814.[109] Z. Yang, Z. Wu, Z. Yin, et al, “Hybrid radar emitter recognition basedon rough k-means classifier and relevance vector machine,”
Sensors ,vol. 13, no. 1, pp. 848-864, 2013.[110] J. Dudczyk, J. Matuszewski, M. Wnuk, “Applying the radiated emissionto the emitter identification,” in
Proc. 15th Int. Conf. on Microwaves,Radar and Wireless Communications , Warsaw, Poland, 2004, pp. 431-434.[111] L. T. Sun, Z. T. Huang, X. Wang, et al, “Overview of radio frequencyfingerprint extraction in specific emitter identification,”
Journal ofRadars , in press. Doi: 10.12000/JR19115. (in Chinese).[112] J. S. Xie, “Robust intra-pulse modulation recognition via sparse repre-sentation,”
IEEE International Radar Conference , 2016, pp. 1-4.[113] Z. Huang, Z. Ma, G. Huang, “Radar waveform recognition based onmultiple autocorrelation Images,,”
IEEE Access , vol. 7, pp. 98653-98668, 2019. Doi: 10.1109/ACCESS.2019.2930250.[114] J. Unden, V. Koivunen, “Automatic radar waveform recognition,”
IEEEJournal of Selected Topics in Signal Processing , vol. 1, no. 1, pp. 124-136, 2007.[115] H. Wang, M. Diao, L. Gao, “Low probability of intercept radarwaveform recognition based on dictionary learning,” in
Proc.2018 10th International Conference on Wireless Communicationsand Signal Processing (WCSP) , Hangzhou, 2018, pp. 1-6,. Doi:10.1109/WCSP.2018.8555906.[116] J. Gao, L. Shen, F. Ye, S. Wang, et al., “Multi-feature radar signalmodulation recognition based on improved PSO algorithm,,”
TheJournal of Engineering , vol. 2019, no. 19, pp. 5588-5592, 2019. Doi:10.1049/joe.2019.0439.[117] W. Gongming, C. Shiwen, H. Xueruobai, et al., “Radar emitter sort-ing and recognition based on time-frequency image union feature,”in
Proc. 2019 IEEE 4th International Conference on Signal andImage Processing (ICSIP) , Wuxi, China, 2019, pp. 165-170. Doi:10.1109/SIPROCESS.2019.8868524.[118] T. Alrubeaan, K. Albagami, A. Ragheb, et al., “An investigation of LPIradar waveforms classification in RoF channels,”
IEEE Access , vol. 7,pp. 124844-124853, 2019. Doi: 10.1109/ACCESS.2019.2938317.[119] V. Iglesias, J. Grajal, P. Royer, et al., “Real-time low-complexityautomatic modulation classifier for pulsed radar signals,”
IEEE Trans-actions on Aerospace and Electronic Systems , vol. 51, no. 1, pp. 108-126, January 2015. Doi: 10.1109/TAES.2014.130183.[120] J. Guo, P. Ge, W. Jin, et al., “Radar signal recognition basedon FCBF and adaBoost algorithm,” in
Proc. 2018 37th ChineseControl Conference (CCC) , Wuhan, 2018, pp. 4185-4190. Doi:10.23919/ChiCC.2018.8483351.[121] G. Revillon, A. Mohammad-Djafari, C. Enderli, “Radar emitters clas-sification and clustering with a scale mixture of normal distributions,”in
Proc. 2018 IEEE Radar Conference (RadarConf18) , Oklahoma City,OK, 2018, pp. 1371-1376. Doi: 10.1109/RADAR.2018.8378764.
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 38 [122] Jun Liu, J. P. Y. Lee, Lingjie Li, et al., “Online clustering algorithmsfor radar emitter classification,”
IEEE Transactions on Pattern Analysisand Machine Intelligence , vol. 27, no. 8, pp. 1185-1196, Aug. 2005.Doi: 10.1109/TPAMI.2005.166.[123] J. Sui, Z. Liu, L. Liu, et al., “Online non-cooperative radar emitterclassification from evolving and imbalanced pulse streams,”
IEEESensors Journal . Doi: 10.1109/JSEN.2020.2981976.[124] Z. Yang, Z. Wu, Z. Yin, et al., “Hybrid radar emitter recognition basedon rough k-means classifier and relevance vector machine,”
Sensors ,vol. 13, no. 1, pp. 848-864, 2013.[125] Y. Zhai, S. Fang, Z. Cheng and X. Ji, “Radar emitter recognitionbased on ambiguity function contour plot,” in
Proc. 2019 IEEE 7thInternational Conference on Computer Science and Network Technol-ogy (ICCSNT) , Dalian, China, 2019, pp. 232-235. Doi: 10.1109/ICC-SNT47585.2019.8962468.[126] S. Ur Rehman, K. Sowerby, C. Coghill, “RF fingerprint extraction fromthe energy envelope of an instantaneous transient signal,” in , Wellington,2012, pp. 90-95. Doi: 10.1109/AusCTW.2012.6164912.[127] Y. Huang, W. Jin, B. Li, et al., “Automatic modulation recogni-tion of radar signals based on Manhattan distance-based features,”
IEEE Access , vol. 7, pp. 41193-41204, 2019. Doi: 10.1109/AC-CESS.2019.2907159.[128] S. Guo, R. E. White, M. Low, “A comparison study of radar emitteridentification based on signal transients,” in
Proc. 2018 IEEE RadarConference (RadarConf18) , Oklahoma City, OK, 2018, pp. 0286-0291.Doi: 10.1109/RADAR.2018.8378572.[129] A. Aubry, A. Bazzoni, V. Carotenuto, et al., “Cumulants-based RadarSpecific Emitter Identification,” in
Proc. 2011 IEEE InternationalWorkshop on Information Forensics and Security , Iguacu Falls, 2011,pp. 1-6. Doi: 10.1109/WIFS.2011.6123155.[130] Z. Liu, Y. Shi, Y. Zeng, et al., “Radar emitter signal detection withconvolutional neural network,” in
Proc. 2019 IEEE 11th InternationalConference on Advanced Infocomm Technology (ICAIT) , Jinan, China,2019, pp. 48-51. Doi: 10.1109/ICAIT.2019.8935926.[131] Z. Qu, C. Hou, C. Hou, et al., “Radar signal intra-pulse modulationrecognition based on convolutional neural network and deep Q-learningnetwork,”
IEEE Access , vol,8, pp.49125 - 4913, 2020.[132] G. Kong, V. Koivunen, “Radar waveform recognition using Fourier-based synchrosqueezing transform and CNN,” in
Proc. 2019 IEEE 8thInternational Workshop on Computational Advances in Multi-SensorAdaptive Processing (CAMSAP) , Le gosier, Guadeloupe, 2019, pp. 664-668. Doi: 10.1109/CAMSAP45676.2019.9022525.[133] E. Yar, M. B. Kocamis, A. Orduyilmaz, “A complete frameworkof radar pulse detection and modulation classification for cognitiveEW,” in
Proc. 2019 27th European Signal Processing Conference(EUSIPCO) , A Coruna, Spain, 2019, pp. 1-5. Doi: 10.23919/EU-SIPCO.2019.8903045.[134] F.C. Akyon, Y. K. Alp, G. Gok, et al., “Deep learning in electronic war-fare systems: Automatic intra-pulse modulation recognition,” in
Proc.2018 26th Signal Processing and Communications Applications Con-ference (SIU) , Izmir, 2018, pp. 1-4. Doi: 10.1109/SIU.2018.8404294.[135] X. Wang, G. Huang, Z. Zhou, et al., “Radar emitter recognition basedon the energy cumulant of short time fourier transform and reinforceddeep belief network,”
Sensors , vol. 18, no. 9, 2018.[136] H. Li, W. Jing, Y. Bai, “Radar emitter recognition based ondeep learning architecture,” in
Proc. 2016 CIE International Con-ference on Radar (RADAR) , Guangzhou, 2016, pp. 1-5. Doi:10.1109/RADAR.2016.8059512.[137] X. Qin, X. Zha, J. Huang, et al, “Radar waveform recogni-tion based on deep residual network,” in
Proc. 2019 IEEE 8thJoint International Information Technology and Artificial IntelligenceConference (ITAIC) , Chongqing, China, 2019, pp. 892-896, doi:10.1109/ITAIC.2019.8785588.[138] H. P. Khanh Nguyen, V. Long Do, Q. T. Dong, “A parallel neuralnetwork-based scheme for radar emitter recognition,” in
Proc. 202014th International Conference on Ubiquitous Information Managementand Communication (IMCOM) , Taichung, Taiwan, 2020, pp. 1-8. Doi:10.1109/IMCOM48794.2020.9001727.[139] F. Wang, C. Yang, S. Huang, et al., “Automatic modulation classifi-cation based on joint feature map and convolutional neural network,,”
IET Radar, Sonar and Navigation , vol. 13, no. 6, pp. 998-1003, Jun.2019. Doi: 10.1049/iet-rsn.2018.5549.[140] A. Dai, H. Zhang, H. Sun, “Automatic modulation classification usingstacked sparse auto-encoders,” in
Proc. 2016 IEEE 13th InternationalConference on Signal Processing (ICSP) , Chengdu, 2016, pp. 248-252.Doi: 10.1109/ICSP.2016.7877834. [141] P. Itkin and N. Levanon, “Ambiguity function based radar waveformclassification and unsupervised adaptation using deep CNN models,” in
Proc. 2019 IEEE International Conference on Microwaves, Antennas,Communications and Electronic Systems (COMCAS) , Tel-Aviv, Israel,2019, pp. 1-6. Doi: 10.1109/COMCAS44984.2019.8958242.[142] X. Li, Z. Liu, Z. Huang, et al., “Radar emitter classification withattention-based multi-RNNs,”
IEEE Communications Letters , doi:10.1109/LCOMM.2020.2995842.[143] X. Li, Z. Liu, Z. Huang, “Attention-based radar PRI modulationrecognition with recurrent neural networks,”
IEEE Access , vol. 8, pp.57426-57436, 2020. Doi: 10.1109/ACCESS.2020.2982654.[144] S. A. Shapero, A. B. Dill, B. O. Odelowo, “Identifying agile waveformswith neural networks,” in [Proc. 2018 21st International Conferenceon Information Fusion (FUSION) , Cambridge, 2018, pp. 745-752. Doi:10.23919/ICIF.2018.8455370.[145] S. Wei, Q. Qu, H. Su, et al., “Intra-pulse modulation radar signal recog-nition based on CLDN network,”
IET Radar, Sonar and Navigation ,vol. 14, no. 6, pp. 803-810, Jun. 2020. Doi: 10.1049/iet-rsn.2019.0436.[146] Y. Shi, H. Ji, “Kernel canonical correlation analysis for specific radaremitter identification,”
Electron. Lett. , vol. 50, no. 18, pp. 1318-1320,2014.[147] G. Huang, Y. Yuan, X. Wang, et al., “Specific emitter identificationbased on nonlinear dynamical characteristics,”
Can. J. Electr. Comput.Eng. , vol. 39, no. 1, pp. 34-41, 2016.[148] J. Dudczyk, A. Kawalec, “Identification of emitter sources in the aspectof their fractal features,”
Bull. Pol. Acad. Sci., Tech. Sci. , vol. 61, no.3, pp. 623-628, 2013.[149] J. Dudczyk, “Radar emission sources identification based on hierar-chical agglomerative clustering for large data sets,”
J. Sens. , pp. 1-9,2016.[150] M. Conning, F. Potgieter, “Analysis of measured radar data for SpecificEmitter Identification,” in
Proc. IEEE radar conference , pp. 35-38,2010.[151] J. Roe, “A review of applications of artificial intelligence techniques tonaval ESM signal processing,” in
Proc. IEE Colloquium on Applicationof Artificial Intelligence Techniques to Signal Processing , London, UK,1989, pp. 5/1-5/5.[152] M. Zhang, M. Diao, L. Gao, and L. Liu, “Neural networks for radarwaveform recognition,”
Symmetry , vol. 9, no. 5, pp. 75, 2017.[153] Z. Qu, W. Wang, C. Hou and C. Hou, “Radar signal intra-Pulsemodulation recognition based on convolutional denoising autoencoderand deep convolutional neural network,”
IEEE Access , vol. 7, pp.112339-112347, 2019.[154] J. Matuszewski, A. Kawalec, “Knowledge-based signal processing forradar identification,” in
Proc. International Conference on ModernProblems of Radio Engineering, Telecommunications and ComputerScience , 2008, pp. 302-305.[155] J. Dudczyk, J. Matuszewski, M. Wnuk, “Applying the relationalmodelling and knowledge-based technologies to the emitter databasedesign,” in
Proc. 14th IEEE International Conference on Microwaves,Radar and Wireless Communications , MIKON, 2002, pp. 172-175.[156] X. Wang, Y. Xiong, B. Tang, et al., “An approach to the modulationrecognition of MIMO radar signals,”
Eurasip Journal on Wireless Com-munications and Networking , 2013. Doi: 10.1186/1687-1499-2013-66.[157] L. H. Nguyen, T. D. Tran. “RFI-radar signal separation via simulta-neous low-rank and sparse recovery,” in
Proc. IEEE radar conference ,2016, pp. 1-5.[158] W. Zhu, M. Li, C. Zeng, et al., “Research on Online Learning ofRadar Emitter Recognition Based on Hull Vector,” in
Proc. IEEEInternational Conference on Data Science in Cyberspace , 2017, pp.328-332.[159] M. Conning, F. Potgieter. “Analysis of measured radar data for SpecificEmitter Identification,” in
Proc. IEEE radar conference , 2010, pp. 35-38.[160] S. T Nguyen, S. Kodituwakku, R. Melino et al., “Signal separationof helicopter radar returns using wavelet-based sparse signal optimiza-tion,” 2016.[161] Li J, Ying Y. “Radar signal recognition algorithm based on entropytheory,” in
Proc. International Conference on Systems , 2014, pp. 718-723.[162] K.P.Behrooz, K.P.Behzad, J.Sciortino, “Automatic Data Sorting UsingNeural Network Techniques,” DTIC Document, 1996.[163] A. W. Ataa, S. N. Abdullah, “Deinterleaving of radar signals and PRFidentification algorithms,”
IET Radar Sonar and Navigation , vol. 1, no.5, pp. 340-347, 2007.
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 39 [164] Z. Liu, P. S. Yu. “Classification, denoising, and deinterleaving ofpulse streams with recurrent neural networks,”
IEEE Transactions onAerospace and Electronic Systems , vol. 55, no. 4, pp. 1624-1639, 2019.[165] Z. Ge, X. Sun, W. Ren, et al, “Improved algorithm of radar pulse repe-tition interval deinterleaving based on pulse correlation,”
IEEE Access ,vol. 7, pp. 30126-30134, 2019. Doi: 10.1109/ACCESS.2019.2901013.[166] O. Torun, M. B. Kocamis, H. Abaci, et al, “Deinterleaving of radarsignals with stagger PRI and dwell-switch PRI types,” ,Antalya, 2017, pp. 1-4, doi: 10.1109/SIU.2017.7960383.[167] H. Mu, J. Gu and Y. Zhao, “A deinterleaving method for mixedpulse signals in complex electromagnetic environment,” , Chengdu, China, 2019, pp. 1-4. Doi: 10.1109/IC-CAIS46528.2019.9074573.[168] A. Erdogan, K. George, “Deinterleaving radar pulse train using neuralnetworks,” , New York, NY, USA,2019, pp. 141-147. Doi: 10.1109/CSE/EUC.2019.00036.[169] X. Li, Z. Liu and Z. Huang, ”Denoising of Radar Pulse Streams WithAutoencoders,” in IEEE Communications Letters, vol. 24, no. 4, pp.797-801, April 2020, doi: 10.1109/LCOMM.2020.2967365.[170] I. Jordanov, N. Petrov, A. Petrozziello , et al., “Supervised radarsignal classification,” in
Proc. International Joint Conference on NeuralNetwork , 2016, pp. 1464-1471.[171] L. Anjaneyulu, N. S. Murthy, N. V. Sarma, et al., “Identification of LPIradar signals by higher order spectra and neural network techniques,”in
Proc. International Conference on Electronic Design , 2008, pp. 1-6.[172] M. R. Azimisadjadi, D. Yao, Q. Huang, et al., “Underwater targetclassification using wavelet packets and neural networks,”
IEEE Trans-actions on Neural Networks , vol. 11, no. 3, pp. 784-794, 2000.[173] Z. Yin, W. Yang, Z. Yang, et al., “A study on radar emitter recognitionbased on spds neural network,”
Information Technology Journal , vol.10, no. 4, pp. 883-888, 2011.[174] Z. Zhang, X. Guan, Y. He, et al., “Study on radar emitter recognitionsignal based on rough sets and RBF neural network,” in
Proc. Inter-national Conference on Machine Learning and Cybernetics , 2009, pp.1225-1230.[175] C. Shieh, C. Lin, “A vector neural network for emitter identification,”
IEEE Transactions on Antennas and Propagation , vol. 50, no. 8, pp.1120-1127, 2002.[176] E. Granger, M. A. Rubin, S. Grossberg, et al., “A what-and-wherefusion neural network for recognition and tracking of multiple radaremitters,”
Neural Networks , vol. 14, no. 3, pp. 325-344, 2001.[177] D. R Pape, J. A Anderson, J. A Carter, et al., “Advanced signalwaveform classifier,” In
Proc. of the 8th Optical Technology forMicrowave Applications , 1997, pp. 162-169.[178] H. Liu, Z. Liu, W. Jiang, et al., “Approach based on combination ofvector neural networks for emitter identification,”
IET Signal Process-ing , vol. 4, no. 2, pp. 137-148, 2010.[179] J. Matuszewski. “Applying the decision trees to radar targets recogni-tion,” in
Proc. International radar symposium , 2010, pp. 1-4.[180] W. Chen, K. Fu, J. Zuo, et al., “Radar emitter classification for largedata set based on weighted-xgboost,”
IET Radar Sonar and Navigation ,vol. 11, no. 8, pp. 1203-1207, 2017.[181] S. U. Rehman , K. W. Sowerby, C. Coghill, et al., “RF fingerprintextraction from the energy envelope of an instantaneous transientsignal,” in
Proc. Australian communications theory workshop , 2002,pp. 90-95.[182] N. Petrov, I. Jordanov, J. Roe, et al., “Radar emitter signals recognitionand classification with feedforward networks,”
Procedia ComputerScience , pp. 1192-1200, 2013.[183] J. Sun, G. Xu, W. Ren, et al., “Radar emitter classification based onunidimensional convolutional neural network,”
IET Radar Sonar andNavigation , vol. 12, no. 8, pp. 862-867, 2018.[184] J. Matuszewski. “The Specific Radar Signature in Electronic Recogni-tion System,”
Przeglad Elektrotechniczny , 2013.[185] D. Zeng, X. Zeng, G. Lu,et al.,“Automatic modulation classificationof radar signals using the generalised time-frequency representation ofzhao, atlas and marks,”
IET radar, sonar and navigation , vol. 5, no. 4,pp. 507-516, 2011.[186] D. Zeng, X. Zeng, H. Cheng, et al., “Automatic modulation classi-fication of radar signals using the rihaczek distribution and houghtransform,”
IET Radar, Sonar and Navigation , vol. 6, no. 5, pp. 322-331, 2012. [187] R. Mingqiu, C. Jinyan, and Z. Yuanqing, “Classification of radarsignals using time-frequency transforms and fuzzy clustering,” in
Proc.IEEE International Conference on In Microwave and Millimeter WaveTechnology (ICMMT) , 2010, pp. 2067-2070.[188] K. Konopko, Y. P. Grishin, and D. Janczak, “Radar signal recognitionbased on time-frequency representations and multidimensional proba-bility density function estimator,” in
Proc. IEEE in Signal ProcessingSymposium (SPSympo) , 2015, pp. 1-6.[189] B. D. Rigling, C. Roush, “Acf-based classification of phase modulatedwaveforms,” in
Proc. IEEE in Radar Conference , 2010, pp. 287-291.[190] C. Wang, H. Gao, and X. Zhang, “Radar signal classificcation based onauto-correlation function and directed graphical model,” in
Proc. IEEEInternational Conference on Signal Processing, Communications andComputing (ICSPCC) , 2016, pp. 1-4.[191] Z. Yu, C. Chen, and W. Jin, “Radar signal automatic classificationbased on pca,” in
Proc. WRI Global Congress on Intelligent Systems ,2009, pp. 216-220.[192] G. Zhang, W. Jin, L. Hu, et al., “Radar emitter signal recognitionbased on support vector machines,” in
Proc. International Conferenceon Control, Automation, Robotics and Vision , 2004, pp. 826-831.[193] Y. Yuan, Z. Huang, Z. Sha, et al., “Spenific emitter identification basedon transient energy tracjectory,”
Progress in Electromagnetics ResearchC , vol. 44, no. 44, pp. 67-82, 2013.[194] J. Xu, M. He, J. Han and C. Chen, “A comprehensive estimationmethod for kernel function of radar signal classifier,”
In ChineseJournal of Electronics , vol. 24, no. 1, pp. 218-222, Jan. 2015.[195] W. Zhu, M. Li, C. Zeng, et al., “Research on online learning of radaremitter recognition based on hull vector,” in
Proc. IEEE InternationalConference on Data Science in Cyberspace , 2017, pp. 328-332.[196] S. Liu, X. Yan, P. Li, et al., “Radar emitter recognition based on SIFTposition and scale features,”
IEEE Transactions on Circuits and SystemsIi-express Briefs , vol. 65, no. 12, pp. 2062-2066, 2018.[197] C. Shieh, C. Lin. “A vector neural network for emitter identification,”
IEEE Transactions on Antennas and Propagation , vol. 50, no. 8, pp.1120-1127, 2002.[198] H. Liu, Z. Liu, W. Jiang, et al., “Approach based on combinationof vector neural networks for emitter identification,”,
IET SignalProcessing , vol. 4, no. 2, pp. 137-148, 2010.[199] Z. Yin, W. Yang, Z. Yang, et al., “A study on radar emitter recognitionbased on spds neural network,”
Information Technology Journal , vol.10, no. 4, pp. 883-888, 2011.[200] Z. Zhang, X. Guan, Y. He, et al., “Study on radar emitter recognitionsignal based on rough sets and RBF neural network,” in
Proc. Inter-national conference on machine learning and cybernetics , 2009, pp.1225-1230.[201] E. Granger, M. A. Rubin, S. Grossberg, et al., “A what-and-wherefusion neural network for recognition and tracking of multiple radaremitters,”
Neural Networks , vol. 14, no. 3, pp. 325-344, 2001.[202] M. D. Wong, A. K. Nandi, “Automatic digital modulation recognitionusing artificial neural network and genetic algorithm,”
Signal Process .vol. 84, pp. 351-365, 2004.[203] J. Matuszewski, “Applying the decision trees to radar targets recogni-tion,” in
Proc. International radar symposium , 2010, pp. 1-4.[204] K. Kim, C. M. Spooner, I. Akbar, et al., “Specific emitter identificationfor cognitive radio with application to IEEE 802.11,” in
Proc. GlobalCommunications Conference , 2008, pp. 1-5.[205] Z. Yang, Z. Wu, Z. Yin, et al., “Hybrid radar emitter recognition basedon rough k-Means classifier and relevance vector machine,”
Sensors ,vol. 13, no. 1, pp. 848-864, 2013.[206] Z. Yang, W. Qiu, H. Sun, et al., “Robust radar emitter recognition basedon the three-dimensional distribution feature and transfer learning,”
Sensors , vol. 16, no. 3, pp. 289-289, 2016.[207] B. M. Pavlyshenko, “Linear, machine learning and probabilistic ap-proaches for time series analysis,” in
Proc. IEEE First Int. Conf. onData Stream Mining and Processing (DSMP) , Lviv, Ukraine, 2016, pp.377-381.[208] L. S. Kim, H. B. Bae, R. M. Kil, et al., “Classification of the trainedand untrained emitter types based on class probability output networks,”
Neurocomputing , pp. 67-75, 2017.[209] J. Guo, P. Ge, W. Jin, et al., “Radar signal recognition based on FCBFand adaBoost algorithm,” in
Proc. Chinese control conference , 2018.[210] J. Sun, G. Xu, W. Ren, et al., “Radar emitter classification based onunidimensional convolutional neural network,”
IET Radar Sonar andNavigation , vol. 12, no. 8, pp. 862-867, 2018.[211] A. Selim, F. Paisana, J. A. Arokkiam, et al., “Spectrum monitoring forradar bands using deep convolutional neural networks,” in
Proc. IEEEGlobal Communications Conference , Singapore, 2017, pp. 1-6.
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 40 [212] “Radar signal recognition based on squeeze-and-excitation networks,”in
Proc. 2019 IEEE Interference on Signal, Data and Signal Processing(ISDSP) , Chongqing, 2019.[213] S. J. Wei, Q. Z. Qu, Y. Wu, et al., “PRI modulation recogni-tion based on squeeze-and-excitation networks,”
IEEE Communi-cations Letters , vol. 24, no. 5, pp. 1047-1051, May 2020. Doi:10.1109/LCOMM.2020.2970397.[214] Q. Qu, S. Wei, Y. Wu, M. Wang, “ACSE networks and autocorrela-tion features for PRI modulation recognition,”
IEEE CommunicationsLetters , 2020. Doi: 10.1109/LCOMM.2020.2992266.[215] S. Wei, Q. Qu, M. Wang, et al., “Automatic modulation recognitionfor radar signals via multi-branch ACSE networks,”
IEEE Access , vol.8, pp. 94923-94935, 2020. Doi: 10.1109/ACCESS.2020.2995203.[216] Z. Liu, P. S. Yu, “Classification, Denoising, and deinterleaving ofpulse treams with recurrent neural networks,”
IEEE Transactions onAerospace and Electronic Systems , vol. 55, no. 4, pp. 1624-1639 2019.[217] M. D. Ming Zhang, L. Guo, “Convolutional neural networks forautomatic cognitive radio waveform recognition,”
IEEE Access , vol.5, pp. 11074-11082, 2017.[218] M. Zhang, M. Diao, L. Gao, L. Liu, “Neural networks for radarwaveform recognition,”
Symmetry , vol. 9, no. 5, pp. 75, 2017.[219] Radar Emitter Signal Detection with Time-frequency image processingand Convolutional Neural Network[220] F. C. Akyon, Y. K. Alp, G. Gok, et al., “Classification of intra-pulsemodulation of radar signals by feature fusion based convolutionalneural networks,” in
Proc. European Signal Processing conference ,2018, PP. 2290-2294.[221] Z. Qu, X. Mao, and Z. Deng, “Radar signal intra-pulse modulationrecognition based on convolutional neural network,”
IEEE Access , vol.6, pp. 43874-43884, 2018.[222] Z. Qu, W. Wang, C. Hou and C. Hou, “Radar signal intra-pulsemodulation recognition based on convolutional denoising autoencoderand deep convolutional neural network,”
IEEE Access , vol. 7, pp.112339-112347, 2019.[223] Z. Liu, X. Mao, Z. Deng, et al., “Radar signal waveform recognitionbased on convolutional denoising autoencoder,” in
Proc. InternationalConference on Communications , 2018, pp. 752-761.[224] Z. Zhou, G. Huang, H. Chen, et al., “Automatic radar waveformrecognition based on deep convolutional denoising auto-encoders,”
Circuits Systems and Signal Processing , vol. 37, no. 9, pp. 4034-4048,2018.[225] J. Wan, X. Yu, Q. Guo, et al., “LPI radar waveform recognition basedon CNN and TPOT,”
Symmetry , vol. 11, no. 5, 2019.[226] Q. Wang, P. Du, J. Yang, et al., “Transferred deep learning basedwaveform recognition for cognitive passive radar,”
Signal Processing ,pp. 259-267, 2019.[227] Q. Guo, X. Yu, G. Ruan, et al., “LPI radar waveform recognition basedon deep convolutional neural network transfer learning,”
Symmetry , vol.11, no. 4, 2019.[228] D. Li, R. Yang, X. Li and S. Zhu, “Radar signal modulation recognitionbased on deep joint learning,”
IEEE Access , vol. 8, pp. 48515-48528,2020. Doi: 10.1109/ACCESS.2020.2978875.[229] X. Wang, G. Huang, Z. Zhou, et al., “Radar emitter recognition basedon the short time fourier transform and convolutional neural networks,”in
Proc. International Congress on Image and Signal Processing , 2017,pp. 1-5.[230] M. Kong, J. Zhang, W. Liu, G. Zhang, “Radar emitter identificationbased on deep convolutional neural network,” in
Proc. 2018 Interna-tional Conference on Control, Automation and Information Sciences(ICCAIS) , Hangzhou, 2018, pp. 309-314.[231] W.Q. Ye and C. Peng, “Recognition algorithm of emitter signals basedon pca + cnn,” in
Proc. 2018 IEEE 3rd Advanced Information Tech-nology, IEEE Electronic and Automation Control Conference (IAEAC) ,2018, pp. 2410-2414.[232] W. Zhang, P. Ge, W. Jin, et al., “Radar signal recognition based onTPOT and LIME,” in
Proc. Chinese Control Conference , 2018.[233] Z. Zhou, J. Zhang, T. Zhang, et al., “A novel algorithm of radaremitter identification based convolutional neural network and randomvector functional-link,” in
Proc. International Conference on NaturalComputation , 2019, PP. 833-842.[234] C. M. Jeong, Y. G. Jung, S. J. Lee, et al., “Deep belief networks basedradar signal classification system,”
Journal of Ambient Intelligence andHumanized Computing , pp. 1-13, 2018.[235] S. Liu, X. Yan, P. Li, et al., “Radar emitter recognition based on SIFTposition and scale features,”
IEEE Transactions on Circuits and SystemsIi-express Briefs , vol. 65, no. 12, pp. 2062-2066, 2018. [236] R. Cao, J. Cao, J. Mei, et al., “Radar emitter identification withbispectrum and hierarchical extreme learning machine,”
MultimediaTools and Applications , vol. 78, no. 20, pp. 28953-28970, 2019.[237] X. Li, Z. Huang, F. Wang, X. Wang, T. Liu, “Toward convolutionalneural networks on pulse repetition interval modulation recognition,”
IEEE Communications Letters , vol. 22, no. 11, pp. 2286-2289, Nov.2018.[238] E. R. Keydel, S. W. Lee, J. T. Moore, “MSTAR extended operatingconditions: A tutorial,” in
Proc. Aerospace/Defense Sensing and Con-trols , 1996, pp. 228-242.[239] H. Wang, S. Chen, F. Xu and Y. Jin, “Application of deep-learningalgorithms to MSTAR data,” in
Proc. 2015 IEEE International Geo-science and Remote Sensing Symposium (IGARSS) , Milan, 2015, pp.3743-3745.[240] Y. Zhou, H. P. Wang, F. Xu, et al., “Polarimetric SAR image classi-fication using deep convolutional neural networks,”
IEEE Geoscienceand Remote Sensing Letters , vol. 13, no. 12, pp. 1935-1939, 2016. Doi:10.1109/LGRS.2016.2618840.[241] F. Xu, H. P. Wang, Y. Q. Jin, “Deep learning as applied in SAR targetrecognition and Terrain classification,”.
Journal of Radars , vol. 6, no.2, pp. 136-148, 2017. Doi: 10.12000/JR16130.(in Chinese)[242] Z. M. Zhang, H. P. Wang, F. Xu, et al., “Complex-valued convolutionalneural network and its application in polarimetric SAR image classi-fication,”
IEEE Transactions on Geoscience and Remote Sensing , vol.55, no. 12, pp. 7177-7188, 2017. Doi: 10.1109/TGRS.2017.2743222.[243] T. Sterling, D. S. Katz, L. Bergman, “High Performance computingsystems for autonomous spaceborne missions,”
International Journalof High Performance Computing Applications , vol. 15, no. 3, pp. 282-296, 2001.[244] http://earth.eo.esa.int/polsarpro/default.html[245] U. A. Force, “MSTAR overview,” [online]. Available:http://tinyurl.com/pc8nh3s, 2013.[246] X. Xing, K. Ji, H. Zou, W. Chen, J. Sun, “Ship classificationin TerraSAR-X images with feature space based sparserepresentation,”
IEEE Geosci. Remote Sens. Lett. , vol.10, no. 6, pp. 1562-1566, Nov. 2013, [online] Available:http://dx.doi.org/10.1109/LGRS.2013.2262073.[247] J. Li, C. Qu, J. Shao, “Ship detection in SAR images based on animproved faster R-CNN,” in
Proc. 2017 SAR in Big Data Era: Models,Methods and Applications (BIGSARDATA) , Beijing, 2017, pp. 1-6.[248] Y. Y. Wang, C. Wang, H. Zhang, et al., “A SAR dataset of ship detectionfor deep learning under complex backgrounds,”
Remote Sensing , vol.11, no. 7, 2019. Doi: 10.3390/rs11070765.[249] S. Xian, Z. R. Wang, Y. R. Sun, et al., “AIR-SARShip-1.0: Highresolution SAR ship detection dataset,”
Journal of Radars , doi:10.12000/JR19097. (in Chinese)[250] S. J. Wei, X. F. Zeng, Q. Z. Qu, et al., “HRSID: A high-resolutionSAR images dataset for ship detection and instance segmentation,”
IEEE Access , doi: 10.1109/ACCESS.2020.300586.[251] Z. Y. Cui, M. R. Zhang, Z. J. Cao, et al., “Image data augmentationfor SAR sensor via generative adversarial nets,”
IEEE Access , vol. 7,pp. 42255-42268, 2019.[252] Y. Ma, L. Yan, W. Y. Zhang, S. Yan, “SAR target recognition basedon transfer learning and data augmentation with LSGANs,” in
Proc.Chinese Automation Congress (CAC) 2019 , 2019, pp. 2334-2337.[253] Z. Wang, L. Du, J. Mao, B. Liu, et al., “SAR target detection based onSSD with data augmentation and transfer learning,”
IEEE Geoscienceand Remote Sensing Letters , vol. 16, no. 1, pp. 150-154, Jan. 2019.[254] J. Y. Lv, Y. Liu, “Data augmentation based on attributed scatteringcenters to train robust CNN for SAR ATR,”
IEEE Access , vol. 7, pp.25459-25473, 2019.[255] Q. Z. Yu, H. B. Hu, X. P. et al., “High-performance SAR automatictarget recognition under limited data condition based on a deep featurefusion network,”
IEEE Access , vol. 7, pp. 165646-165658, 2019.[256] Z. L. Huang, Z. X. Pan, B. Lei, “Transfer learning with deep convolu-tional neural network for SAR target classification with limited labeleddata,”
Remote Sensing , vol. 9, pp. 907, 2017.[257] Y. Yan, “Convolutional neural networks based on augmented trainingsamples for synthetic aperture radar target recognition,”
Journal ofElectronic Imaging , vol. 27, pp. 1, 2018.[258] J. Ding, B. Chen, H. Liu et al., “Convolutional neural network withdata augmentation for SAR target recognition,”
IEEE Geoscience andRemote Sensing Letters , vol. 13, no. 3, pp. 364-368, March 2016.[259] H. Wang, S. Chen, F. Xu et al., “Application of deep-learning algo-rithms to MSTAR data,” in
Proc. 2015 IEEE International Geoscienceand Remote Sensing Symposium (IGARSS) , Milan, 2015, pp. 3743-3745.
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 41 [260] M. Everingham, S. M. Ali Eslami, L. V. Gool, “The PASCAL visualobject classes challenge: A retrospective,”
International Journal ofComputer Vision , vol. 111, pp. 98-136, 2015.[261] L. Q. Huang, B. Liu, B. Y. Li, et al., “OpenSARShip: A datasetdedicated to sentinel-1 ship interpretation,”
IEEE Journal of SelectedTopics in Applied Earth Observations and Remote Sensing , vol. 11, no.1, pp. 195-208, 2018.[262] C. Oliver, S. Quegan, “Understanding synthetic aperture radar images,”Norwood, MA, USA: Artech House, 1998.[263] A. Baraldi, F. Parmiggiani, “A refined gamma MAP SAR speckle filterwith improved geometrical adaptivity,”
IEEE Trans. Geosci. RemoteSens. , vol. 33, no. 5, pp. 1245-1257, Sep. 1995.[264] V. S. Frost, J. A. Stiles, K. S. Shanmugan, et al., “A model for radarimages and its application to adaptive digital filtering of multiplicativenoise,”
IEEE Trans. Pattern Anal. Mach. Intell. , vol. PAMI-4, no. 2,pp. 157-166, Mar. 1982.[265] K. Dabov, A. Foi, V. Katkovnik, et al., “Image denoising with block-matching and 3D filtering,” in
Proc. SPIE , vol. 6064, 2006.[266] S. Parrilli, M. Poderico, C. V. Angelino, “A nonlocal SAR imagedenoising algorithm based on LLMMSE wavelet shrinkage,”
IEEETrans. Geosci. Remote Sens. , vol. 50, no. 2, pp. 606-616, Feb. 2012.[267] A. Achim, P. Tsakalides, A. Bezerianos, “SAR image denoising viaBayesian wavelet shrinkage based on heavy-tailed modeling,”
IEEETrans. Geosci. Remote Sens. , vol. 41, no. 8, pp. 1773-1784, Aug. 2003.[268] F. Argenti, L. Alparone, “Speckle removal from SAR images in theundecimated wavelet domain,”
IEEE Trans. Geosci. Remote Sens. , vol.40, no. 11, pp. 2363-2374, 2002.[269] H. Xie, L. E. Pierce, F. T. Ulaby, “SAR speckle reduction using waveletdenoising and Markov random field modeling,”
IEEE Trans. Geosci.Remote Sens. , vol. 40, no. 10, pp. 2196-2212, 2002.[270] V. M. Patel, G. R. Easley, R. Chellappa, et al., “Separated component-based restoration of speckled SAR images,”
IEEE Trans. Geosci.Remote Sens. , vol. 52, no. 2, pp. 1019-1029, Feb. 2014.[271] F. Argenti, A. Lapini, T. Bianchi, et al., “A tutorial on speckle reductionin synthetic aperture radar images,”
IEEE Geosci. Remote Sens. Mag. ,vol. 1, no. 3, pp. 6-35, Sep. 2013.[272] D. Mishra, S. Tyagi, S. Chaudhury, et al, “Despeckling CNN withensembles of classical outputs,” in
Proc. 2018 24th InternationalConference on Pattern Recognition (ICPR) , 2018, pp. 3802-3807.[273] G. Chierchia, D. Cozzolino, G. Poggi, et al., “SAR image despecklingthrough convolutional neural networks,” in
Proc. 2017 IEEE Interna-tional Geoscience and Remote Sensing Symposium , Fort Worth, TX,USA, 2017, pp. 5438-5441.[274] J. P. Zhao, W. W. Guo, B. Liu, et al., “Convolutional neural network-based SAR image classification with noisy labels,”
Journal of Radars ,vol. 6, no. 5, pp. 514-523, 2017. Doi: 10.12000/JR16140.(in Chinese)[275] P. Y. Wang, H. Zhang, V. M. Patel, “SAR image despeckling usinga convolutional neural network,”
IEEE Signal Processing Letters , vol.24, no. 12, pp. 1763-1767, 2017. Doi: 10.1109/LSP.2017.2758203.[276] J. Wang, T. Zheng, P. Lei, et al., “Ground target classification in noisySAR images using convolutional neural networks,”
IEEE Journal ofSelected Topics in Applied Earth Observations and Remote Sensing ,vol. 11, no. 11, pp. 4180-4192, Nov. 2018.[277] H. Parikh, S. Patel, V. Patel, “Analysis of denoising techniques forspeckle noise removal in synthetic aperture radar images,” in
Proc.2018 International Conference on Advances in Computing Communi-cations and Informatics (ICACCI) , 2018, pp. 671-677.[278] G. Di Martino, M. Poderico, G. Poggi, et al., “Benchmarking frame-work for SAR despeckling,”
IEEE Trans. Geosci. Remote Sens. , vol.52, no. 3, pp. 1596-1615, Mar. 2014.[279] S. Mukherjee, A. Zimmer, N. K. Kottayil, et al., “CNN-Based InSARdenoising and coherence metric,” , 2018, pp. 1-4.[280] N. Brodu, “Low-rankness transfer for denoising Sentinel-1 SAR im-ages,” in
Proc. 2018 9th International Symposium on Signal ImageVideo and Communications (ISIVC) , 2018, pp. 106-111.[281] S. Vitale, G. Ferraioli, V. Pascazio, “A new ratio image based CNN al-gorithm for SAR despeckling,” in
Proc. IEEE International Geoscienceand Remote Sensing Symposium IGARSS , 2019, pp. 9494-9497.[282] G. Ferraioli, V. Pascazio, S. Vitale, “A novel cost function for despeck-ling using convolutional neural networks,” in
Proc. Remote SensingEvent (JURSE) , Joint Urban, 2019, pp. 1-4.[283] R. Farhadiani, S. Homayouni, A. Safari, “Hybrid SAR speckle reduc-tion using complex wavelet shrinkage and non-local PCA-based Ffilter-ing,”
IEEE Journal of Selected Topics in Applied Earth Observationsand Remote Sensing , vol. 12, no. 5, pp. 1489-1496, 2019. [284] Z. H. Zhou, E. Y. Lam, C. Lee, “Nonlocal means filtering based speckleremoval utilizing the maximum a posteriori estimation and the totalvariation image prior,”
IEEE Access , vol. 7, pp. 99231-99243, 2019.[285] X. L. Yang, L. Denis, F. Tupin, et. al, “SAR image despeckling usingpre-trained convolutional neural network models,” in
Proc. RemoteSensing Event (JURSE) , Joint Urban, 2019, pp. 1-4.[286] K. Ravani, S. Saboo, J. S. Bhatt, “A practical approach for SAR imagedespeckling using deep learning,” , 2019, pp. 2957-2960.[287] O. Ronneberger et al., “U-net: Convolutional networks for biomedicalimage segmentation,” in
Proc. International Conference on MedicalImage Computing and Computer-Assisted Intervention , 2015, pp. 234-241.[288] D. Cozzolino, L. Verdoliva, G. Scarpa, et al., “Nonlocal SAR imagedespeckling by convolutional neural networks,” in , 2019,pp. 5117-5120.[289] S. Vitale, D. Cozzolino, G. Scarpa, et al., “Guided patchwise nonlocalSAR despeckling,”
IEEE Transactions on Geoscience and RemoteSensing , vol. 57, no. 9, pp. 6484-6498, 2019.[290] F. Tupin, L. Denis, C. A. Deledalle, et al., “Ten years of patch-based approaches for sar imaging: A review”, in
Proc. 2019 IEEEInternational Geoscience and Remote Sensing Symposium (IGARSS) ,2019, pp. 5105-5108.[291] F. Gu, H. Zhang, C. Wang, “A two-component deep learning networkfor SAR Iimage denoising,”
IEEE Access , vol. 8, pp. 17792-17803,2020.[292] Q. Zhang, Q. Yuan, J. Li, et al., “Learning a dilated residual networkfor SAR image despeckling,”
Remote Sens. , vol. 10, no. 2, pp. 196,Jan. 2018.[293] S. Liu, T. Liu, L. Gao, et al., “Convolutional neural network and guidedfiltering for SAR image denoising,”
Remote Sens. , vol. 11, no. 6, pp.702, Mar. 2019.[294] S. Foucher, M. Beaulieu, M. Dahmane, et al., “Deep speckle noisefiltering,” in
Proc. IEEE Int. Geosci. Remote Sens. Symp. (IGARSS) ,2017, pp. 5311-5314.[295] F. Gu, H. Zhang, C. Wang, et al., “Residual encoder-decoder networkintroduced for multisource SAR image despeckling,” in
Proc. SARin Big Data Era: Models Methods and Applications (BIGSARDATA) ,2017, pp. 1-5.[296] X. Tang, L. Zhang, X. Ding, “SAR image despeckling with a multilayerperceptron neural network,
Int. J. Digit. Earth , vol. 12, no. 3, pp. 354-374, Mar. 2019.[297] H. Wang, S. Chen, F. Xu et al., “Application of deep-learning algo-rithms to MSTAR data,” in
Proc. 2015 IEEE International Geoscienceand Remote Sensing Symposium (IGARSS) , Milan, 2015, pp. 3743-3745.[298] Y. B. Bai, C. Gao, S. Singh, et al., “A framework of rapidregional tsunami damage recognition from post-event TerraSAR-X imagery using deep neural networks,”
IEEE Geoscience andRemote Sensing Letters , vol. 15, no. 1, pp. 43-47, 2018. Doi:10.1109/LGRS.2017.2772349.[299] Y. Zhou, H. P. Wang, F. Xu, et al., “Polarimetric SAR image classi-fication using deep convolutional neural networks,”
IEEE Geoscienceand Remote Sensing Letters , vol. 13, no. 12, pp. 1935-1939, 2016. Doi:10.1109/LGRS.2016.2618840.[300] F. Xu, H. P. Wang, Y. Q. Jin, “Deep learning as applied in SAR targetrecognition and Terrain classification,”
Journal of Radars , vol. 6, no.2, pp. 136-148, 2017. Doi: 10.12000/JR16130.(in Chinese)[301] Z. W. Zhang, H. P. Wang, F. Xu, et al., “Complex-valued convolutionalneural network and its application in polarimetric SAR image classi-fication,”
IEEE Transactions on Geoscience and Remote Sensing , vol.55, no. 12, pp. 7177-7188, 2017. Doi: 10.1109/TGRS.2017.2743222.[302] A. G. Mullissa, C. Persello, A. Stein, “PolSARNet: A deep fullyconvolutional network for polarimetric SAR image classification,”
IEEE Journal of Selected Topics in Applied Earth Observations andRemote Sensing , vol. 12, no. 12, pp. 5300-5309, Dec. 2019. Doi:10.1109/JSTARS.2019.2956650.[303] R. Shang, G. Wang, M A Okoth, et al., “Complex-valued convolutionalautoencoder and spatial pixel-squares refinement for polarimetric SARimage classification,”
Remote Sensing , vol. 11, no. 5, 2019.[304] Z. Y. Cui, M. G. Zhang, Z. J. Cao, “Image data augmentation forSAR sensor via generative adversarial nets,”
IEEE Access , vol. 7, pp.42255-42268, 2019.[305] Y. Ma, Y. Liang, W. Y. Zhang, “SAR target recognition based ontransfer learning and data augmentation with LSGANs,” in
Proc.Chinese Automation Congress (CAC) , 2019, pp. 2334-2337.
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 42 [306] Z. Wang, L. Du, J. Mao, et al., “SAR target detection based on SSDwith data augmentation and transfer learning,”
IEEE Geoscience andRemote Sensing Letters , vol. 16, no. 1, pp. 150-154, Jan. 2019.[307] J. Y. Lv, Y. Liu, “Data augmentation based on attributed scatteringcenters to train robust CNN for SAR ATR”,
IEEE Access , vol. 7, pp.25459-25473, 2019.[308] Q. Z. Yu, H. B. Hu, X. P. Geng, et al., “High-performance SARautomatic target recognition under limited data condition based on adeep feature fusion network,”
IEEE Access , vol. 7, pp. 165646-165658,2019.[309] R. H. Shang, J. M. Wang, L. C. Jiao, et al., “SAR targets clas-sification based on deep memory convolution neural networks andtransfer parameters,”
IEEE Journal of Selected Topics in Applied EarthObservations and Remote Sensing , vol. 11, no. 8, pp. 2834-2846, 2018.[310] J. Ding, B. Chen, H. Liu, et al., “Convolutional neural network withdata augmentation for SAR target recognition,”
IEEE Geoscience andRemote Sensing Letters , vol. 13, no. 3, pp. 364-368, March 2016.[311] H. Furukawa, “Deep learning for target Classification from SARimagery data augmentation and translation invariance,”
ArXiv :1708.07920, 2017.[312] Z. Zhao, et al., “Discriminant deep belief network for high-resolution SAR image classification,”
Pattern Recognition , 2016.http://dx.doi.org/10.1016/j.patcog.2016.05.028.[313] S. Chen, H. Wang, “SAR target recognition based on deep learning,”in
Proc. International Conference Data Science , 2014, pp. 541-547.[314] Q. Zhao, J. C. Principe, “ Support vector machines for SAR automatictarget recognition,”
IEEE Transactions on Aerospace and ElectronicSystems , vol. 37, no. 2, pp. 643-654, 2001.[315] C. Georgios, “SVM-based target recognition from synthetic apertureradar images using target region outline descriptors,”
Nonlinear Anal.Theory Meth. Appl. , vol. 71, no. 12, pp. e2934-e2939, Dec. 2009.[online] Available: http://dx.doi.org/10.1016/j.na.2009.07.030.[316] J. Zhao. C. Principe, “Support vector machines for SAR au-tomatic target recognition,”
IEEE Trans. Aerosp. Electron. Syst. ,vol. 37, no. 2, pp. 643-654, Apr. 2001. [online] Available:http://dx.doi.org/10.1109/7.937475.[317] S. A. Stanhope, J. M. Daida, “Genetic programming for automatic tar-get classification and recognition in synthetic aperture radar imagery,”
Evol. Programming VII , Berlin, Germany, Springer, pp. 735-744, 1998.[online] Available: http://dx.doi.org/10.1007/BFb0040824.[318] Y. Sun, Z. Liu, S. Todorovic, et al., “Adaptive boosting forSAR automatic target recognition,”
IEEE Trans. Aerosp. Electron.Syst. , vol. 43, no. 1, pp. 112-125, Jan. 2007. [online] Available:http://dx.doi.org/10.1109/TAES.2007.357120.[319] T. Perciano, F. Tupin, R. Hirata, et al., “A hierarchical Markovrandom field for road network extraction and its application withoptical and SAR data,” in
Proc. IEEE Int. Geosci. Remote Sens.Symp. (IGARSS) , Jul. 2011, pp. 1159-1162, [online] Available:http://dx.doi.org/10.1109/IGARSS.2011.6049403.[320] R. Weisenseel, W. Karl, D. Castanon, et al., “Markov ran-dom field segmentation methods for SAR target chips,” in
Proc.SPIE , Aug. 1999, vol. 3721, pp. 462-473, [online] Available:http://dx.doi.org/10.1117/12.357662.[321] D. A. Clausi, “Improved texture recognition of SAR sea iceimagery by data fusion of MRF features with traditionalmethods,” in
Proc. IEEE Int. Geosci. Remote Sens. Symp.(IGARSS) , Jul. 2001, vol. 3, pp. 1170-1172, [online] Available:http://dx.doi.org/10.1109/IGARSS.2001.976781.[322] H. Ruohong, Y. Ruliang, “SAR target recognition based on MRF andgabor wavelet feature extraction,” in
Proc. IEEE Int. Geosci. RemoteSens. Symp. (IGARSS) , Jul. 2008, vol. 2, pp. II-907-II-910, , [online]Available: http://dx.doi.org/10.1109/IGARSS.2008.4779142.[323] Z. Lin, K. F. Ji, M. Kang, et al., “Deep convolutional highway unitnetwork for SAR target classification with limited labeled trainingdata,”
IEEE Geoscience and Remote Sensing Letters , vol. 14, no. 7,pp. 1091-1095, 2017.[324] I. M. Gorovyi, D. S. Sharapov, “Comparative analysis of convolutionalneural networks and support vector machines for automatic targetrecognition,” in
Proc. IEEE Microwaves Radar and Remote SensingSymposium (MRRS) , 2017, pp. 63-66.[325] J. F. Pei, Y. L. Huang, W. B. Huo, “SAR automatic target recognitionbased on multiview deep learning framework,”
IEEE Transactions onGeoscience and Remote Sensing , vol. 56, no. 4, pp. 2196-2210, 2018.[326] F. Zhang, C. Hu, Q. Yin, “Multi-aspect-aware bidirectional LSTMnetworks for synthetic aperture radar target recognition,”
IEEE Access ,vol. 5, pp. 26880-26891, 2017. [327] B. Y. Ding, G. J. Wen, X. H. Huang, et. al, “Data augmentation bymultilevel reconstruction using attributed scattering center for SARtarget recognition,”
IEEE Geoscience and Remote Sensing Letters , vol.14, no. 6, pp. 979-983, 2017.[328] J. F. Pei, Y. L. Huang, W. B. Huo, “Multi-view SAR ATR basedon networks ensemble and graph search,” in
Proc. 2018 IEEE RadarConference (RadarConf18) , 2018, pp. 0355-0360.[329] Z. L. Huang, Z. X. Pan, B. Lei, “Transfer Learning with deepconvolutional neural network for SAR target classification with limitedlabeled data,”
Remote Sensing , vol. 9, pp. 907, 2017.[330] B. Y. Ding, G. J. Wen, C. H. Ma, “An efficient and robust frameworkfor SAR target recognition by hierarchically fusing global and localfeatures,”
IEEE Transactions on Image Processing , vol. 27, no. 12, pp.5983-5995, 2018.[331] J. F. Pei, W. B. Huo, Q. H. Zhang, et al., “Multi-view bistatic syntheticaperture radar target recognition based on multi-input deep convolu-tional neural network,” in
Proc. 2018 IEEE International Geoscienceand Remote Sensing Symposium (IGARSS) , 2018, pp. 2314-2317[332] J. H. Cho, C. G. Park, “Multiple feature aggregation using con-volutional neural networks for SAR image-based automatic targetrecognition,”
IEEE Geoscience and Remote Sensing Letters , vol. 15,no. 12, pp. 1882-1886, 2018.[333] W. Qi, G. J. Wen, “SAR target classification method based on con-volutional neural network,” in
Proc. 2018 IEEE 9th InternationalConference on Software Engineering and Service Science (ICSESS) , 2018, pp. 1151-1155.[334] X. R. Bai, R. H. Xue, L. Wang, et al., “Sequence SAR imageclassification based on bidirectional convolution-recurrent network,”
IEEE Transactions on Geoscience and Remote Sensing , vol. 57, no.11, pp. 9223-9235, 2019.[335] S. Zhang, Z. D. Wen, Z. G. Liu, et al., “Rotation awareness based self-supervised learning for SAR target recognition,” in
Proc. 2019 IEEEInternational Geoscience and Remote Sensing Symposium (IGARSS)
IEEEAccess , vol. 7, pp. 165127-165142, 2019.[337] C. Ning, W. B. Liu, G. Zhang, et al., “Synthetic aperture radar targetrecognition using weighted multi-task kernel sparse representation,”
IEEE Access , vol. 7, pp. 181202-181212, 2019.[338] Y. Gu, Y. Xu, J. Liu, “SAR ATR by decision fusion of multiple randomconvolution features,” in
Proc. 2019 22th International Conference onInformation Fusion , 2019, pp. 1-8.[339] M. Kang, K. F. Ji, X. G. Leng, et al., “Synthetic aperture radartarget recognition with feature fusion based on a stacked autoencoder,”
Sensors , vol. 17, pp. 192, 2017.[340] X. Z. Zhang, Y. J. Wang, Z. Y. Tan, et al., “Two-stage multi-taskrepresentation learning for synthetic aperture radar (SAR) target imagesclassification,”
Sensors , vol. 17, pp. 2506, 2017.[341] J. Q. Shao, C. W. Qu, J. W. Li, et al., “A lightweight convolutionalneural network based on visual attention for SAR image target classi-fication,”
Sensors , vol. 18, pp. 3039, 2018.[342] S. Chen, H. Wang, F. Xu, et al., “Target classification using the deepconvolutional networks for SAR images,”
IEEE Trans. Geosci. RemoteSens. , vol. 54, pp. 4806-4817, 2016.[343] P. F. Zhao, K. Liu, H. Zou, et al, “Multi-stream convolutional neuralnetwork for SAR automatic target recognition,”
Remote Sensing , vol.10, pp. 1473, 2018.[344] M. Touafria, Q. Yang, “A concurrent and hierarchy target learningarchitecture for classification in SAR application,”
Sensors , vol. 18,pp. 3218, 2018.[345] Z. Z. Tian, L. P. Wang, R. H. Zhan, et al., “Classification via weightedkernel CNN: application to SAR target recognition,”
InternationalJournal of Remote Sensing , pp. 1, 2018.[346] G. Q. Huang, X. G. Liu, J. P. Hui, et al., “A novel group squeezeexcitation sparsely connected convolutional networks for SAR targetclassification,”
International Journal of Remote Sensing , pp. 1, 2019.[347] J. Y. Lv, “Exploiting multi-level deep features via joint sparse rep-resentation with application to SAR target recognition,”
InternationalJournal of Remote Sensing , pp. 1, 2019.[348] B. Q. Qi, H. T. Jing, H. Chen, et al., “Target recognition in syntheticaperture radar image based on PCANet,”
The Journal of Engineering ,vol. 2019, no. 21, pp. 7309-7312, 2019.[349] Z. Ren, B. Hou, Z. Wen, et al., “Patch-sorted deep feature learning forhigh resolution SAR Image Classification,”
IEEE Journal of SelectedTopics in Applied Earth Observations and Remote Sensing , vol. 11, no.9, pp. 3113-3126, Sept. 2018. Doi: 10.1109/JSTARS.2018.2851023.
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 43 [350] H. Liu, S. Yang, S. Gou, et al., “Polarimetric SAR feature extractionwith neighborhood preservation-based deep learning,”
IEEE Journalof Selected Topics in Applied Earth Observations and Remote Sens-ing , vol. 10, no. 4, pp. 1456-1466, April 2017. Doi: 10.1109/JS-TARS.2016.2618891.[351] Y. Zhai et al., “MF-SarNet: Effective CNN with data augmentation forSAR automatic target recognition,”
The Journal of Engineering , vol.2019, no. 19, pp. 5813-5818, 10 2019. Doi: 10.1049/joe.2019.0218.[352] D. Guo, B. Chen, “SAR image target recognition via deep Bayesiangenerative network,” in
Proc. 2017 International Workshop on RemoteSensing with Intelligent Processing (RSIP) , Shanghai, 2017, pp. 1-4.Doi: 10.1109/RSIP.2017.7958814.[353] T. Yan, X. Sun, W. Yang, et al., “Unsupervised classification of polari-metric SAR images using deep embedding network,” in
Proc. 2017 SARin Big Data Era: Models, Methods and Applications (BIGSARDATA) ,Beijing, 2017, pp. 1-6, doi: 10.1109/BIGSARDATA.2017.8124939.[354] W. Wu, H. Li, L. Zhang, et al., “High-resolution PolSAR sceneclassification with pretrained deep convnets and manifold polarimetricparameters,”
IEEE Transactions on Geoscience and Remote Sensing ,vol. 56, no. 10, pp. 6159-6168, 2018.[355] D. Song et al., “A novel marine oil spillage identification schemebased on convolution neural network feature extraction from fullypolarimetric SAR imagery,”
IEEE Access , vol. 8, pp. 59801-59820,2020. Doi: 10.1109/ACCESS.2020.2979219.[356] T. Gadhiya, A. K. Roy, “Optimized Wishart network for an efficientclassification of multifrequency PolSAR data,”
IEEE Geoscience andRemote Sensing Letters , vol. 15, no. 11, pp. 1720-1724, Nov. 2018.Doi: 10.1109/LGRS.2018.2861081.[357] S. A. Wagner, “SAR ATR by a combination of convolutional neural net-work and support vector machines,”
IEEE Transactions on Aerospaceand Electronic Systems , vol. 52, no. 6, pp. 2861-2872, December 2016.Doi: 10.1109/TAES.2016.160061.[358] F. Gao, T. Huang, J. Wang, et al., “Combining deep convolutional neu-ral network and SVM to SAR image target recognition,” in
Proc. 2017IEEE International Conference on Internet of Things (iThings) andIEEE Green Computing and Communications (GreenCom) and IEEECyber, Physical and Social Computing (CPSCom) and IEEE SmartData (SmartData) , Exeter, 2017, pp. 1082-1085. Doi: 10.1109/iThings-GreenCom-CPSCom-SmartData.2017.165.[359] S. R. Tian, C. Wang, H. Zhang, “Hierarchical feature exttratction forobject recogition in complex SAR image using modified convolutionalauto-encoder,” in
Proc. 2017 IEEE International Geoscience andRemote Sensing Symposium (IGARSS) , Fort Worth, TX, 2017, pp. 854-857. Doi: 10.1109/IGARSS.2017.8127087.[360] F. Gao, T. Huang, J. Sun, et al., “A new algorithm of SAR image targetrecognition based on improved deep convolutional neural network,”
Cognitive Computation , vol. 11, no. 6, pp. 809-824, 2019.[361] M. A. Mufti, E. A. Hadhrami, B. Taha, et al., “Using transfer learningtechnique for SAR automatic target recognition,” in
Proc. SPIE 11197,SPIE Future Sensing Technologies, 111970A , 12 November 2019,https://doi.org/10.1117/12.2538012.[362] X. Chen, X. Peng, R. Duan, et al., “Deep kernel learning method forSAR image target recognition,”
Review of Scientific Instruments , vol.88, no. 10, 2017.[363] Y. Cho, S. Shin, S. Yim, et al., “Multistage fusion with dissimilarityregularization for SAR/IR target recognition,”
IEEE Access , vol. 7, pp.728-740, 2019. Doi: 10.1109/ACCESS.2018.2885736.[364] H. Liu, F. Shang, S. Yang, et al., “Sparse manifold-regularizedneural networks for polarimetric SAR terrain classification,”
IEEETransactions on Neural Networks and Learning Systems , doi:10.1109/TNNLS.2019.2935027.[365] R. Xue, X. Bai, F. Zhou, “Spatial-temporal ensemble convolution forsequence SAR target classification,”
IEEE Transactions on Geoscienceand Remote Sensing , doi: 10.1109/TGRS.2020.2997288.[366] W. Liu, S. Gou, W. Chen, et al., “Classification of interferometricsynthetic aperture radar image with deep learning approach,” in
Proc.2016 CIE International Conference on Radar (RADAR) , Guangzhou,2016, pp. 1-3. Doi: 10.1109/RADAR.2016.8059200.[367] J. Oh, G. Youm, M. Kim, “SPAM-Net: A CNN-based SAR targetrecognition network with pose angle marginalization learning,”
IEEETransactions on Circuits and Systems for Video Technology , doi:10.1109/TCSVT.2020.2987346.[368] L. Gu, F. He, S. Yang, “Crop classification based on deep learning innortheast China using SAR and optical imagery,” in
Proc. 2019 SARin Big Data Era (BIGSARDATA) , Beijing, China, 2019, pp. 1-4. Doi:10.1109/BIGSARDATA.2019.8858437. [369] J. Geng, H. Wang, J. Fan et al., “Classification of fusing SAR and multi-spectral image via deep bimodal autoencoders,” in
Proc. 2017 IEEE In-ternational Geoscience and Remote Sensing Symposium (IGARSS) , FortWorth, TX, 2017, pp. 823-826. Doi: 10.1109/IGARSS.2017.8127079.[370] F. Gao, T. Huang, J. Wang, et al., “Dual-branch deep convolution neuralnetwork for polarimetric SAR image classification,”
Applied Sciences ,vol. 7, no. 5, 2017.[371] H. Dong, L. Zhang, B. Zou, “Densely connected convolutional neu-ral network based polarimetric SAR image classification,” in
Proc.IGARSS 2019 - 2019 IEEE International Geoscience and RemoteSensing Symposium , Yokohama, Japan, 2019, pp. 3764-3767. Doi:10.1109/IGARSS.2019.8900292.[372] C. Wang et al., “Automated geophysical classification of sentinel-1wave mode SAR images through deep-learning,” in
Proc. IGARSS 2018- 2018 IEEE International Geoscience and Remote Sensing Symposium ,Valencia, 2018, pp. 1776-1779. Doi: 10.1109/IGARSS.2018.8518354.[373] Y. Guo, E. Chen, Z. Li, et al., “Convolutional highway unit network forlarge-scale classification with GF-3 dual-PolSar data,” in
IGARSS 2018- 2018 IEEE International Geoscience and Remote Sensing Symposium ,Valencia, 2018, pp. 2424-2427. Doi: 10.1109/IGARSS.2018.8518737.[374] R. Xue, X. Bai, “2D-temporal convolution for target recognition ofSAR sequence image,” in
Proc. 2019 6th Asia-Pacific Conference onSynthetic Aperture Radar (APSAR) , Xiamen, China, 2019, pp. 1-4, doi:10.1109/APSAR46974.2019.9048438.[375] L. Wang, X. Xu, H. Dong, et al., “Exploring convolutional lstm forPolsar image classification,” in
Proc. IGARSS 2018 - 2018 IEEEInternational Geoscience and Remote Sensing Symposium , Valencia,2018, pp. 8452-8455. Doi: 10.1109/IGARSS.2018.8518517.[376] L. Li, L. Ma, L. Jiao, et al., “Complex contourlet-CNN for polarimetricSAR image classification,”
Pattern Recognition , 2019.[377] Z. Huang, M. Datcu, Z. Pan, et al., “Deep SAR-Net: Learning objectsfrom signals,”
Isprs Journal of Photogrammetry and Remote Sensing ,pp. 179-193, 2020.[378] C. He, M. Tu, D. Xiong, et al., “Nonlinear manifold learning integratedwith fully convolutional networks for PolSAR image classification,”
Remote Sensing , vol. 12, no. 4, 2020.[379] J. Zhao, M. Datcu, Z. Zhang, et al., “Contrastive-regulated CNN inthe complex domain: A method to learn physical scattering signaturesfrom flexible PolSAR images,”
IEEE Transactions on Geoscience andRemote Sensing , vol. 57, no. 12, pp. 10116-10135, 2019.[380] X. Wang, Z. Cao, Z. Cui, et al., “PolSAR image classification basedon deep polarimetric feature and contextual information,”
Journal ofApplied Remote Sensing , vol. 13, no. 03, 2019.[381] P. Zhang, L. Chen, Z. Li, et al., “Automatic extraction of water andshadow from SAR images based on a multi-resolution dense encoderand decoder network,”
Sensors , vol. 19, no. 16, 2019.[382] L. Chen, X. Cui, Z. Li, et al., “A new deep learning algorithm for SARscene classification based on spatial statistical modeling and featuresre-calibration,”
Sensors , vol. 19, no. 11, 2019.[383] Z. Zhao, L. Guo, M. Jia, et al., “The generalized gamma-DBN forhigh-resolution SAR image classification,”
Remote Sensing , vol. 10,no. 6, 2018.[384] Y. Chen, L. Jiao, Y. Li, et al., “Multilayer projective dictionary pairlearning and sparse autoencoder for PolSAR image classification,”
IEEE Transactions on Geoscience and Remote Sensing , vol. 55, no.12, pp. 6683-6694, 2017.[385] L. J. Peng, X. H. Liu, M. Liu, et al., “SAR target recognition andposture estimation using spatial pyramid pooling within CNN,” in
Proc.2017 International Conference on Optical Instruments and Technology:Optoelectronic Imaging/Spectroscopy and Signal Processing Technol-ogy , 2018. Doi: 10.1117/12.2285913.[386] E. Ndikumana, D. H. Minh, N. Baghdadi, et al., “Deep recurrentneural network for agricultural classification using multitemporal SARSentinel-1 for Camargue, France,”
Remote Sensing , vol. 10, no. 8, 2018.[387] W. Zhang, Y. Zhu, Q. Fu, “Semi-supervised deep transfer learning-Based on adversarial feature learning for label limited SAR targetrecognition,”
IEEE Access , vol. 7, pp. 152412-152420, 2019. Doi:10.1109/ACCESS.2019.2948404.[388] K. Huang, W. Nie, N. Luo, “Fully polarized SAR imagery classificationbased on deep reinforcement learning method using multiple polari-metric features,”
IEEE Journal of Selected Topics in Applied EarthObservations and Remote Sensing , vol. 12, no. 10, pp. 3719-3730,Oct. 2019. Doi: 10.1109/JSTARS.2019.2913445.[389] W. Zhang, Y. Zhu and Q. Fu, “Adversarial deep domain adaptationfor multi-band SAR images classification,”
IEEE Access , vol. 7, pp.78571-78583, 2019. Doi: 10.1109/ACCESS.2019.2922844.
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 44 [390] C. Zheng, X. Jiang, X. Liu, “Semi-supervised SAR ATR viamulti-discriminator generative adversarial network,”
IEEE SensorsJournal , vol. 19, no. 17, pp. 7525-7533, 1 Sept.1, 2019. Doi:10.1109/JSEN.2019.2915379.[391] K. Wang, G. Zhang, Y. Leng, et al., “Synthetic aperture radar imagegeneration with deep generative models,”
IEEE Geoscience and Re-mote Sensing Letters , vol. 16, no. 6, pp. 912-916, June 2019. Doi:10.1109/LGRS.2018.2884898.[392] Q. Song, F. Xu, “Zero-shot learning of SAR target feature spacewith deep generative neural networks,”
IEEE Geoscience and RemoteSensing Letters , vol. 14, no. 12, pp. 2245-2249, Dec. 2017. Doi:10.1109/LGRS.2017.2758900.[393] F. Ye, W. Luo, M. Dong, et al., “SAR image retrieval based onunsupervised domain adaptation and clustering,”
IEEE Geoscience andRemote Sensing Letters , vol. 16, no. 9, pp. 1482-1486, Sept. 2019. Doi:10.1109/LGRS.2019.2896948.[394] S. Deng, L. Du, C. Li, et al., “SAR automatic target recognition basedon euclidean distance restricted autoencoder,”
IEEE Journal of SelectedTopics in Applied Earth Observations and Remote Sensing , vol. 10, no.7, pp. 3323-3333, July 2017. Doi: 10.1109/JSTARS.2017.2670083.[395] S. Tian, C. Wang, H. Zhang, et al., “SAR object classification usingthe DAE with a modified triplet restriction,”
IET Radar, Sonar andNavigation , vol. 13, no. 7, pp. 1081-1091, 7 2019. Doi: 10.1049/iet-rsn.2018.5413.[396] J. Tang, F. Zhang, Q. Yin, et al., “Synthetic aperture radar targetrecognition of incomplete training datasets via Siamese network,”
TheJournal of Engineering , vol. 2019, no. 20, pp. 6845-6847, 10 2019.Doi: 10.1049/joe.2019.0566.[397] H. Zhu, N. Lin, H. Leung, et al., “Target classification from SARimagery based on the pixel grayscale decline by graph convolutionalneural network,”
IEEE Sensors Letters , vol. 4, no. 6, pp. 1-4, June2020. Doi: 10.1109/LSENS.2020.2995060.[398] F. Zhang, Y. Wang, J. Ni, et al., “SAR target small sample recognitionbased on CNN cascaded features and adaBoost rotation forest,”
IEEEGeoscience and Remote Sensing Letters , vol. 17, no. 6, pp. 1008-1012,June 2020. Doi: 10.1109/LGRS.2019.2939156.[399] T. Jiang, Z. Cui, Z. Zhou, et al., “Data augmentation with Gaborfilter in deep convolutional neural networks for sar target recogni-tion,” in
Proc. IGARSS 2018 - 2018 IEEE International Geoscienceand Remote Sensing Symposium , Valencia, 2018, pp. 689-692. Doi:10.1109/IGARSS.2018.8518792.[400] Q. Song, F. Xu Y. Jin, “SAR image representation learning withadversarial autoencoder networks,” in
Proc. IGARSS 2019 - 2019 IEEEInternational Geoscience and Remote Sensing Symposium , Yokohama,Japan, 2019, pp. 9498-9501, doi: 10.1109/IGARSS.2019.8898922.[401] H. Bi et al., “An active deep learning approach for minimally-supervised polsar image classification,” in
Proc. IGARSS 2019- 2019 IEEE International Geoscience and Remote SensingSymposium , Yokohama, Japan, 2019, pp. 3185-3188, doi:10.1109/IGARSS.2019.8899214.[402] S. Tian, C. Wang, H. Zhang, “SAR object classification with a multi-scale convolutional auto-encoder,” in
Proc. 2019 SAR in Big Data Era(BIGSARDATA) , Beijing, China, 2019, pp. 1-4. Doi: 10.1109/BIGSAR-DATA.2019.8858491.[403] W. Hua, S. Wang, W. Xie, et al., “Dual-channel convolutional neu-ral network for polarimetric SAR images classification,” in
Proc.IGARSS 2019 - 2019 IEEE International Geoscience and RemoteSensing Symposium , Yokohama, Japan, 2019, pp. 3201-3204. Doi:10.1109/IGARSS.2019.8899103.[404] Y. Zhang, X. Sun, H. Sun, et al., “High resolution SAR im-age classification with deeper convolutional neural network,” in
Proc. IGARSS 2018 - 2018 IEEE International Geoscience andRemote Sensing Symposium , Valencia, 2018, pp. 2374-2377. Doi:10.1109/IGARSS.2018.8518829.[405] C. Wang et al., “A deep unsupervised learning method for SARobject classification,” in
Proc. 2019 6th Asia-Pacific Conference onSynthetic Aperture Radar (APSAR) , Xiamen, China, 2019, pp. 1-5, doi:10.1109/APSAR46974.2019.9048377.[406] M. Rostami, S. Kolouri, E. Eaton, et al., “SAR image classifica-tion using few-shot cross-domain transfer learning,,” in
Proc. 2019IEEE/CVF Conference on Computer Vision and Pattern RecognitionWorkshops (CVPRW) , Long Beach, CA, USA, 2019, pp. 907-915. Doi:10.1109/CVPRW.2019.00120.[407] L. Wang, X. Bai, F. Zhou, “Few-shot SAR ATR based on Conv-BiLSTM prototypical networks,” in
Proc. 2019 6th Asia-Pacific Con-ference on Synthetic Aperture Radar (APSAR) , Xiamen, China, 2019,pp. 1-5. Doi: 10.1109/APSAR46974.2019.9048492. [408] C. Liu, W. Liao, H. Li, et al., “A deep-neural-network-based hybridmethod for semi-supervised classification of polarimetric SAR data,”in
Proc. 2019 6th Asia-Pacific Conference on Synthetic ApertureRadar (APSAR) , Xiamen, China, 2019, pp. 1-5. Doi: 10.1109/AP-SAR46974.2019.9048529.[409] Y. K. Zhai, W. B. Deng, Y. Xu, et al., “Robust SAR automatic targetrecognition based on transferred MS-CNN with L2-Regularization,” in
Computational Intelligence and Neuroscience , vol. 2019, no. 3, pp.1-13, November 2019.[410] Y, Hou, Y, Bai, T. Xu, et al., “Deep convolutional neural networkstructural design for synthetic aperture radar image target recognitionbased on incomplete training data and displacement insensitivity,”
Journal of Electronic Imaging , vol. 28, no. 5, 2019.[411] M. Ahishali, S. Kiranyaz, T. Ince, et al., ”Dual and single polar-ized SAR image classification using compact convolutional neuralnetworks,”
Remote Sensing , vol. 11, no. 11, 2019.[412] Q. Rui, F. Xiongjun. D. Jian, et al., “A semi-greedy neuralnetwork CAE-HL-CNN for SAR target recognition with limitedtraining data,”
International Journal of Remote Sensing , 2020.doi:10.1080/01431161.2020.1766149.[413] Wang R, Wang Y., “Classification of PolSAR image using neural nonlo-cal stacked sparse autoencoders with virtual adversarial regularization,”
Remote Sensing , vol. 11, no. 9, 2019.[414] F. Gao, F. Ma F, J. Wang, et al., “Semi-supervised generative adversarialnets with multiple generators for SAR image recognition,”
Sensors , vol.18, no. 8, 2018.[415] Gao F, Yang Y, Wang J, et al., “A deep convolutional generativeadversarial networks (DCGANs)-based semi-supervised method forobject recognition in synthetic aperture radar (SAR) images,”
RemoteSensing , vol. 10, no. 6, 2018.[416] S. Chen, C. Tao, X. Wang, S. Xiao, “Polarimetric SAR tar-gets detection and classification with deep convolutional neuralnetwork,” in
Proc. 2018 Progress in Electromagnetics ResearchSymposium (PIERS-Toyama) , Toyama, 2018, pp. 2227-2234. Doi:10.23919/PIERS.2018.8597856.[417] H. Zhu, W. Wang, R. Leung, “SAR target classification based on radarimage luminance analysis by deep learning,”
IEEE Sensors Letters , vol.4, no. 3, pp. 1-4, March 2020. Doi: 10.1109/LSENS.2020.2976836.[418] J. Geng, J. Fan, H. Wang, et al., “High-resolution SAR image classi-fication via deep convolutional autoencoders,”
IEEE Geoscience andRemote Sensing Letters , vol. 12, no. 11, pp. 2351-2355, Nov. 2015.Doi: 10.1109/LGRS.2015.2478256.[419] J. Geng, H. Wang, J. Fan, et al., “Deep supervised and contractiveneural network for SAR image classification,”
IEEE Transactions onGeoscience and Remote Sensing , vol. 55, no. 4, pp. 2442-2459, April2017. Doi: 10.1109/TGRS.2016.2645226.[420] J. Geng, X. Ma, J. Fan, et al., “Semisupervised classification of po-larimetric SAR image via superpixel restrained deep neural network,”
IEEE Geoscience and Remote Sensing Letters , vol. 15, no. 1, pp. 122-126, Jan. 2018. Doi: 10.1109/LGRS.2017.2777450.[421] Y. Kwak, W. Song, S. Kim, “Speckle-noise-invariant convolutionalneural network for SAR target recognition,”
IEEE Geoscience andRemote Sensing Letters , vol. 16, no. 4, pp. 549-553, April 2019. Doi:10.1109/LGRS.2018.2877599.[422] Z. Huang, C. O. Dumitru, Z. Pan, et al., “Classificationof large-scale high-resolution SAR images with deep transferlearning,”
IEEE Geoscience and Remote Sensing Letters . Doi:10.1109/LGRS.2020.2965558.[423] Y. Zhou et al., “Complex background SAR target recognition based onconvolution neural network,” in
Proc. 2019 6th Asia-Pacific Conferenceon Synthetic Aperture Radar (APSAR) , Xiamen, China, 2019, pp. 1-4.Doi: 10.1109/APSAR46974.2019.9048279.[424] N. Inkawhich, E. Davis, U. Majumder, et al., “Advanced techniques forrobust SAR ATR: Mitigating noise and phase errors,” in
Proc. 2020IEEE International Radar Conference (RADAR) , Washington, DC,USA, 2020, pp. 844-849. Doi: 10.1109/RADAR42522.2020.9114784.[425] C. Li, L. Du, S. Deng, et al., “Point-wise discriminative auto-encoderwith application on robust radar automatic target recognition,”
SignalProcessing , 2020.[426] L. Zhang, W. Ma, D. Zhang, et al., “Stacked sparse autoencoderin PolSAR data classification using local spatial information,”
IEEEGeoscience and Remote Sensing Letters , vol. 13, no. 9, pp. 1359-1363,2016.[427] Y. Jiang, Z. Chi, “A scale-invariant framework for image classificationwith deep learning,” in
Proc. 2017 IEEE International Conference onSystems, Man, and Cybernetics (SMC) , Banff, AB, 2017, pp. 1019-1024. Doi: 10.1109/SMC.2017.8122744.
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 45 [428] T. Huang, Q. X. Zhang, J. B. Liu, et al., “Adversarial attacks on deep-learning-based SAR image target recognition,”
Journal of Network andComputer Applications , vol. 162, no. 15, July, 2020.[429] R. J Soldin, D. N Macdonald, M. D Reisman, et al., “HySARNet: Ahybrid machine learning approach to synthetic aperture radar automatictarget recognition,” in
Proc. SPIE 10988, Automatic Target RecognitionXXIX , May 2019, https://doi.org/10.1117/12.2518155.[430] X. Chen, J. Deng, “A robust polarmetric SAR terrain classificationbased on sparse deep autoencoder model combined with waveletkernel-based classifier,,”
IEEE Access , vol. 8, pp. 64810-64819, 2020.Doi: 10.1109/ACCESS.2020.2983478.[431] J. Geng, X. Deng, X. Ma, “Transfer learning for SAR im-age classification via deep joint distribution adaptation networks,”
IEEE Transactions on Geoscience and Remote Sensing , doi:10.1109/TGRS.2020.2964679.[432] J. Fan, X. Wang, X. Wang, et, al., “Incremental Wishart broad learningsystem for fast PolSAR image classification,”
IEEE Geoscience andRemote Sensing Letters , vol. 16, no. 12, pp. 1854-1858, Dec. 2019.Doi: 10.1109/LGRS.2019.2913999.[433] T. Gadhiya, A. K. Roy, “Superpixel-driven optimized Wishart networkfor fast PolSAR image classification using global k -means algorithm,” IEEE Transactions on Geoscience and Remote Sensing , vol. 58, no. 1,pp. 97-109, Jan. 2020. Doi: 10.1109/TGRS.2019.2933483.[434] Z. Wang, X. Xu, “Efficient deep convolutional neural networksusing CReLU for ATR with limited SAR images,”
The Journalof Engineering , vol. 2019, no. 21, pp. 7615-7618, 11 2019. Doi:10.1049/joe.2019.0567.[435] H. Liu, S. Yang, S. Gou, et al., “Fast classification for large polarimetricSAR data based on refined spatial-anchor graph,”
IEEE Geoscience andRemote Sensing Letters , vol. 14, no. 9, pp. 1589-1593, Sept. 2017. Doi:10.1109/LGRS.2017.2724844.[436] H. Zhu, N. Lin, H. Leung, et al., “Target classification from SARimagery based on the pixel grayscale decline by graph convolutionalneural network,”
IEEE Sensors Letters , vol. 4, no. 6, pp. 1-4, June2020. Doi: 10.1109/LSENS.2020.2995060.[437] D. Li, Y. Gu, S. Gou, et al., “Full polarization SAR imageclassification using deep learning with shallow feature,” in
Proc.2017 IEEE International Geoscience and Remote Sensing Sym-posium (IGARSS) , Fort Worth, TX, 2017, pp. 4566-4569. Doi:10.1109/IGARSS.2017.8128018.[438] C. Zhao, P. Wang, J. Wang, et al., “A maritime target detector based onCNN and embedded device for GF-3 images,” in
Proc. 2019 6th Asia-Pacific Conference on Synthetic Aperture Radar (APSAR) , Xiamen,China, 2019, pp. 1-4. Doi: 10.1109/APSAR46974.2019.9048264.[439] S. Wagner, K. Barth, S. Brggenwirth, “A deep learning SAR ATRsystem using regularization and prioritized classes,” in
Proc. 2017 IEEERadar Conference (RadarConf) , Seattle, WA, 2017, pp. 0772-0777.Doi: 10.1109/RADAR.2017.7944307.[440] Z. Ying, C. Xuan, Y. Zhai, et al., “TAI-SARNET: Deep transferredatrous-inception CNN for small samples SAR ATR,”
Sensors , vol. 20,no. 6, 2020.[441] C. Ozcan, O. K .Ersoy, I. U. Ogul, et al., ”Fast texture classification ofdenoised SAR image patches using GLCM on spark,
Turkish Journalof Electrical Engineering and Computer Sciences , vol. 28, no. 1, pp.182-195, 2020.[442] Y. Hou, Y. Bai, T. Xu, et al., “Deep convolutional neural networkstructural design for synthetic aperture radar image target recognitionbased on incomplete training data and displacement insensitivity,”
Journal of Electronic Imaging , vol. 28, no. 5, 2019.[443] H. Syed, R. Bryla, U. K. Majumder, et al., “Semi-random deep neuralnetworks for near real-time target classification,”
Algorithms for Syn-thetic Aperture Radar Imagery XXVI , 2019. Doi: 10.1117/12.2520237.[444] R. Min, H. Lan, Z. Cao, et al., “A gradually distilled CNN for SARtarget recognition,”
IEEE Access , pp. 42190-42200, 2019.[445] H. Chen, F. Zhang, B. Tang, et al., “Slim and efficient neural net-work design for resource-constrained SAR target recognition,”
RemoteSensing , vol. 10, no. 10, 2018.[446] J. H. Zhang, H. J. Song, B. B. Zhou, “SAR target classification basedon deep forest model,”
Remote Sensing , vol. 12, pp. 128, 2020.[447] Y. Xue, J. F. Pei, Y. L. Huang, et al., “Target recognition for SARimages based on heterogeneous CNN ensemble,” , 2018, pp. 0507-0512.[448] C. Coman, R. Thaens, “A deep learning SAR target classificationexperiment on MSTAR dataset,” in
Proc. 2018 19th InternationalRadar Symposium (IRS) , pp. 1-6, 2018. [449] O. Kechagias-Stamatis, N. Aouf, “Fusing deep learning and sparsecoding for SAR ATR,”
IEEE Transactions on Aerospace and ElectronicSystems , vol. 55, no. 2, pp. 785-797, 2019.[450] S. Kazemi, B. Yonel, B. Yazici, “Deep learning for direct automatictarget recognition from SAR Data,” in
Proc. 2019 IEEE Radar Con-ference (RadarConf) , pp. 1-6, 2019.[451] J. J. Liu, X. J. Fu, K. Q. Liu, et al., “Spotlight SAR image recognitionbased on dual-channel feature map convolutional neural network,” in
Proc. 2019 IEEE 4th International Conference on Signal and ImageProcessing (ICSIP) , pp. 65-69, 2019.[452] Y.B. Dong, H. Zhang, C. Wang, Y. Y. Wang, “Fine-grained shipclassification based on deep residual learning for high-resolution SARimages,”
Remote Sensing Letters , vol. 10, pp. 1095, 2019.[453] G. Gao, L. Liu, L. Zhao, et al., “An adaptive and fast CFAR algorithmbased on automatic censoring for target detection in high-resolutionSAR images,”
IEEE Trans. Geosci. Remote Sens. , vol. 47, pp. 1685-1697, 2008.[454] A. Farrouki, M. Barkat, “Automatic censoring CFAR detector based onordered data variability for nonhomogeneous environments,” in
Proc.IEE Proc.-Radar Sonar Navig. , 2005, pp. 43-51.[455] Huang, X.; Yang, W.; Zhang, H.; Xia, G.S. Automatic ship detectionin SAR images using multi-scale heterogeneities and an acontrariodecision. Remote Sens. 2015, 7, 7695-7711.[456] X. Xing, K. Ji, H. Zou, “Ship classification in TerraSAR-X images withfeature space based sparse representation,”
IEEE Geosci. Remote Sens.Lett. , vol. 10, no. 6, pp. 1562-1566, Nov. 2013, [online] Available:http://dx.doi.org/10.1109/LGRS.2013.2262073.[457] F. Ma, F. Gao, J. Wang, et al., “A novel biologically inspired targetdetection method based on saliency analysis in SAR imagery,”
Neuro-computing , 2019.[458] T. Xie, W. Zhang, L. Yang, “Inshore ship detection based on level setmethod and visual saliency for SAR images,”
Sensors (Basel) , vol. 18,no. 11, pp. 3877, 2018. Doi:10.3390/s18113877.[459] R. Girshick, J. Donahue, T. Darrell, “Rich feature hierarchies foraccurate object detection and semantic segmentation,” in
Proc. ofthe IEEE Conference on Computer Vision and Pattern Recognition ,Columbus, OH, USA, 23-28 June 2014, pp. 580-587.[460] R. Girshick, “Fast r-cnn,” in
Proc. of the IEEE International Conferenceon Computer Vision , Santiago, Chile, 7-13 Dec. 2015, pp. 1440-1448.[461] S. Ren, K. He, R. Girshick, “Faster R-CNN: Towards real-time objectdetection with region proposal networks,”
IEEE Transactions on Pat-tern Analysis and Machine Intelligence , vol. 39, no. 6, pp. 1137-1149,2016.[462] K. He, G. Gkioxari, P. Dollr, “Mask r-cnn.
In Proceedings of the IEEEInternational Conference on Computer Vision , Venice, Italy, 22-29 Oct.2017, pp. 2961-2969.[463] Z. Cai, N. Vasconcelos, “Cascade r-cnn: Delving into high qualityobject detection,”
In Proceedings of the IEEE Conference on ComputerVision and Pattern Recognition , Salt Lake City, UT, USA, 18-23 June2018, pp. 6154-6162.[464] T. Y. Lin, P. Dollr, R. Girshick, “Feature pyramid networks for objectdetection,”
In Proceedings of the IEEE Conference on Computer Visionand Pattern Recognition , Honolulu, HI, USA, 21-26 July 2017, pp.2117-2125.[465] J. Redmon, S. Divvala, R. Girshick, “ You only look once: Unified,real-time object detection,”
In Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition , Las Vegas, NV, USA, 27-30June 2016, pp. 779-788.[466] J. Redmon, A. Farhadi, “YOLO9000: Better, faster, stronger,”
InProceedings of the IEEE Conference on Computer Vision and PatternRecognition , Honolulu, HI, USA, 21-26 July 2017, pp. 7263-7271.[467] J. Redmon, A. Farhadi, “Yolov3: An incremental improvement,”
ArXiv :1804.02767, 2018.[468] B. Alexey, W. Chien-Yao, L. H. Y. Mark, “YOLOv4: Optimal speedand accuracy of object detection,”
ArXiv : 2004.10934, 2020.[469] P. Hurtik, V. Molek, J. Hula, et al., “Poly-YOLO: higher speed, moreprecise detection and instance segmentation for YOLOv3,” arXiv :2005.13243, 2020.[470] W. Liu, D. Anguelov, D. Erhan, et al., “Ssd: Single shot multibox de-tector,”
In Proceedings of the 14th European Conference on ComputerVision , Amsterdam, The Netherlands, 8-16 October 2016, pp. 21-37.[471] T. Y. Lin, P. Goyal, R. Girshick, He, “Focal loss for dense objectdetection,”
In Proceedings of the IEEE International Conference onComputer Vision , Venice, Italy, 22-29 October 2017, pp. 2980-2988.[472] J. Li, C. Qu, J. Shao, “Ship detection in SAR images based on animproved faster R-CNN,” in
Proc.2017 SAR in Big Data Era: Models,Methods and Applications (BIGSARDATA) , Beijing, 2017, pp. 1-6.
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 46 [473] Q. Fan, F. Chen, M. Cheng, et al., “Ship detection using a fully con-volutional network with compact polarimetric SAR images,”
RemoteSensing , vol. 11, no. 18, 2019.[474] F. Gao, W. Shi, J. Wang, et al., “Enhanced feature extraction forship detection from multi-resolution and multi-scene synthetic apertureradar (SAR) images,”
Remote Sens. , vol. 11, no. 22, 2019.[475] F. Ma, F. Gao, J. Wang, et al., “A novel biologically inspired targetdetection method based on saliency analysis in SAR imagery,”
Neuro-computing , 2019.[476] M. Kang, K. Ji, X. Leng, et al., “Contextual region-based convolutionalneural network with multilayer fusion for SAR ship detection,”
RemoteSensing , vol. 9, no. 8, 2017.[477] Q. An, Z. Pan, H. You. “Ship detection in Gaofen-3 SAR imagesbased on sea clutter distribution analysis and deep convolutionalneural network,”
Sensors (Basel) , vol. 18, no. 2, pp. 334, 2018.Doi:10.3390/s18020334.[478] Y. Y. Wang, C. Wang, H. Zhang, “ Combining a single shot multiboxdetector with transfer learning for ship detection using sentinel-1 SARimages,”
Remote Sensing Letters , vol. 9, no. 8, pp. 780-788, 2018. Doi:10.1080/2150704X.2018.1475770.[479] T. Xie, W. Zhang, L. Yang, “Inshore ship detection based on level setmethod and visual saliency for SAR images,”
Sensors (Basel) , vol. 18,no. 11, 2018. Doi:10.3390/s18113877.[480] Y. Y. Wang, C. Wang, H. Zhang, et al., “Automatic ship detectionbased on RetinaNet using multi-resolution Gaofen-3 imagery,”
RemoteSensing , vol. 11, no. 5, 2019.[481] J. Jiao, Y. Zhang, H. Sun, “A densely connected end-to-end neuralnetwork for multiscale and multiscene SAR ship detection,”
IEEEAccess , vol. 6, pp. 20881-20892, 2018.[482] T. Zheng, J. Wang, P. Lei, “Deep learning based target detection methodwith multi-features in SAR imagery,” in
Proc. 2019 6th Asia-PacificConference on Synthetic Aperture Radar (APSAR) , Xiamen, China,2019, pp. 1-4. Doi: 10.1109/APSAR46974.2019.9048509.[483] H. M. Khan, Y. Z. Cai, “Ship detection in SAR Image using YOLOv2,”in
Proc. 2018 37th Chinese Control Conference (CCC) , 2018, pp. 9495-9499.[484] Y. C. Gui, X. H. Li, L. Xue, “A scale transfer convolution network forsmall ship detection in SAR images,” in
Proc. 2019 IEEE 8th JointInternational on Information Technology and Artificial IntelligenceConference (ITAIC) , 2019, pp. 1845-1849.[485] C. Chen, C. He, C. H. Hu, “A deep neural network based on anattention mechanism for SAR ship detection in multiscale and complexscenarios,”
IEEE Access , vol. 7, pp. 104848-104863, 2019.[486] X. H. Zhang, H. P. Wang, C. G. Xu, “A lightweight feature optimizingnetwork for ship detection in SAR image,”
IEEE Access , vol. 7, pp.141662-141678, 2019.[487] Z. Y. Cui, Q. Li, Z. J. Cao, “Dense attention pyramid networks formulti-scale ship detection in SAR images,”
IEEE Transactions onGeoscience and Remote Sensing , vol. 57, no. 11, pp. 8983-8997, 2019.[488] C. Chen, C. He, C. H. Hu, et al., “MSARN: A deep neural networkbased on an adaptive recalibration mechanism for multiscale andarbitrary-oriented SAR ship detection,”
IEEE Access , vol. 7, pp.159262-159283, 2019.[489] N. Y. Liu, Z. Y. Cui, Z. J. Cao, “Scale-transferrable pyramid networkfor multi-scale ship detection in Sar images,” in
Proc. 2019 IEEEInternational Geoscience and Remote Sensing Symposium (IGARSS) ,2019, pp. 1-4.[490] Q. Li, R. Min, Z. Y. Cui, “Multiscale ship detection based on denseattention pyramid network in Sar images,” in
Proc. 2019 IEEE Inter-national Geoscience and Remote Sensing Symposium (IGARSS) , 2019,pp. 5-8.[491] X. Nie, M. Y. Duan, H. X. Ding, “Attention mask R-CNN for shipdetection and segmentation from remote sensing images,”
IEEE Access ,vol. 8, pp. 9325-9334, 2020.[492] S. J. Wei, H. Su, J. Ming,et al., “Precise and robust ship detection forhigh-resolution SAR imagery based on HR-SDNet,”
Remote Sensing ,vol. 12, pp. 167, 2020.[493] T. Zhang, X. Zhang, “High-speed ship detection in SAR images basedon a grid convolutional neural network,”
Remote Sensing , vol. 11, no.1206, 2019.[494] Y. Duan, F. Liu, L. Jiao, et al., “SAR image segmentationbased on convolutional-wavelet neural network and markov ran-dom field,”
Pattern Recognition , vol. 64, pp. 255-267, 2017.https://doi.org/10.1016/j.patcog.2016.11.015.[495] T. Wu, X. Ruan, X. Chen, et al., “A modified method for theestimation of SAR target aspect angle based on MRF segmentation,” in
Proc. SPIE , Oct. 2009, vol. 7495, pp. 74. [online] Available:http://dx.doi.org/10.1117/12.830709.[496] B. Y. Ding, G. J. Wen, C. H. Ma, et al., “Evaluation of target seg-mentation on SAR target recognition,” in
Proc. 2017 4th InternationalConference on Information Cybernetics and Computational SocialSystems (ICCSS) , 2017, pp. 663-667.[497] X. Gao, D. Qin, J. Gao, et al., ”Resolution enhancement for inverse syn-thetic aperture radar images using a deep residual network,”
Microwaveand Optical Technology Letters , vol. 62. no. 4, pp. 1588-1593, 2020.[498] J. Mitchell, S. Tjuatja, “ISAR imaging in the presence of quasi-random multiplicative noise using convolutional deep learning,” in
Proc. International Geoscience and Remote Sensing Symposium , 2019,pp. 2583-2586.[499] S. Zhang, Y. Liu, X. Li, “Fast sparse aperture ISAR autofocusingand imaging via ADMM based sparse Bayesian learning,”
IEEETransactions on Image Processing , vol. 29, pp. 3213-3226, 2020. Doi:10.1109/TIP.2019.2957939.[500] X. He, N. Tong, X. Hu, et al., “High-resolution ISAR imaging of fastrotating targets based on pattern-coupled Bayesian strategy for multiplemeasurement vectors,”
Digital Signal Processing , pp.151-159, 2019.[501] S. Zhang, Y. Liu, X. Li, et al., “Joint sparse aperture ISAR autofocusingand scaling via modified newton method-based variational Bayesianinference,”
IEEE Transactions on Geoscience and Remote Sensing , vol.57, no. 7, pp. 4857-4869, 2019.[502] B. Xue, N. Tong, “DIOD: Fast and efficient weakly semi-superviseddeep complex ISAR object detection,”
IEEE Transactions on Systems,Man, and Cybernetics , vol. 49, no. 11, pp. 3991-4003, 2019.[503] Xue B, Tong N, Xu X, et al. DIOD: Fast, Semi-Supervised Deep ISARObject Detection[J]. IEEE Sensors Journal, 2019, 19(3): 1073-1081.[504] H. Yang, Y. Zhang, W. Ding, “Multiple heterogeneous P-DCNNsensemble with stacking algorithm: A novel recognition method ofspace target ISAR images under the condition of small sample set,”
IEEE Access , vol. 8, pp. 75543-75570, 2020. Doi: 10.1109/AC-CESS.2020.2989162.[505] B. Xue, N. Tong, “Real-world ISAR object recognition and relationdiscovery using deep relation graph learning,”
IEEE Access , vol. 7, pp.43906-43914, 2019. Doi: 10.1109/ACCESS.2019.2896293.[506] S. Zaied, A. Toumi, A. Khenchaf, et al., “Target classification usingconvolutional deep learning and auto-encoder models,” in
Proc. Inter-national Conference on Advanced Technologies for Signal and ImageProcessing , 2018, pp. 1-6.[507] K. D. Uttecht, C. X. Chen, J. C. Dickinson, et al., “A comparison ofmachine learning methods for target recognition using ISAR imagery,”in
Proc. Proceedings of SPIE , 2011.[508] X. He, N. Tong, X. Hu, “Automatic recognition of ISAR im-ages based on deep learning,” in
Proc. 2016 CIE InternationalConference on Radar (RADAR) , Guangzhou, 2016, pp. 1-4. Doi:10.1109/RADAR.2016.8059230.[509] B. Li, H. Li, W. B. Huang, “Modified PPCA methods for radar HRRProbust recognition,”
Journal of Northwestern Polytechnical University , vol. 34, no. 6, 2016.(in Chinese)[510] J. Yuan, W. Liu, G. Zhang, “Application of dictionary learning algo-rithm in HRRP based on statistical modeling,”
XI Tong Gong ChengYu Dian Zi Ji Shu/systems Engineering & Electronics , vol. 40, no. 4,pp. 762-767, 2018.[511] D. L. Xu, L. Du, P. H. Wang, et al., “Radar HRRP target recognitionby utilizing multitask sparse learning with a small training data size,”
Journal of Xidian University , vol. 43, no. 2, pp. 23-28, 2016.(inChinese)[512] L. Li, Z. Liu, “Noise-robust multi-feature joint learning HRRP recog-nition method,”
Journal of Xidian University vol, 45, pp. 57-62, 2018.[513] Y. Guo, H. Xiao, Y. Kan, et al., “Learning using privileged informationfor HRRP-based radar target recognition,”
IET Signal Processing vol.12, no. 2, pp. 188-197, 2017.[514] M. Li, G. Zhou, B. Zhao, et al., “Sparse representation denoisingfor radar high resolution range profiling,”
International Journal ofAntennas and Propagation , pp. 1-8, 2014.[515] J. N. Wu, Y. G. Chen, D. J. Feng, et al., “Target recog-nition for polarimetric HRRP based on pre-classification andmodel matching,”
Systems Engineering and Electronics
IET Radar Sonar and Navigation , vol. 12. no. 4, pp. 417-425,2018.[517] M. Pan, J. Jiang, Q. Kong, et al., “Radar HRRP target recognition basedon t-SNE segmentation and discriminant deep belief network,”
IEEE
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 47
Geoscience and Remote Sensing Letters , vol. 14, no. 9, pp. 1609-1613,2017.[518] X. Peng, X. Gao, Y. Zhang, et al., “An adaptive feature learning modelfor sequential radar high resolution range profile recognition,”
Sensors ,vol. 17, no. 7, 2017.[519] K. Liao, J. Si, F. Zhu, et al., “Radar HRRP target recognition based onconcatenated deep neural networks,”
IEEE Access , pp. 29211-29218,2018.[520] F. Zhao, Y. Liu, K. Huo, et al., “Radar HRRP target recognition basedon stacked autoencoder and extreme learning machine,”
Sensors , vol.18, no. 1, 2018.[521] Y. Jiang, Y. Li, J. Cai, et al., “Robust automatic target recognition viaHRRP sequence based on scatterer matching,”
Sensors , vol. 18, no. 2,2018. Doi: 10.3390/s18020593.[522] L. Du, H. He, L. Zhao, et al., “Noise robust radar HRRP targetrecognition based on scatterer matching algorithm,”
IEEE SensorsJournal , vol. 16, no. 6, pp. 1743-1753, March15, 2016.[523] L. Li, Z. Liu, “Noise-robust HRRP target recognition method viasparse-low-rank representation,”
Electronics Letters , vol. 53, no. 24,pp. 1602-1604, 2017.[524] W. Dai, G. Zhang, Y. Zhang, et al., “HRRP classification based onmulti-scale fusion sparsity preserving projections,”
Electronics Letters ,vol. 53, no. 11, pp. 748-750, 2017.[525] J. Liu, N. Fang, Y. J. Xie, et al., “Scale-space theory-based multi-scalefeatures for aircraft classification using HRRP,”
Electronics Letters ,vol. 52, no. 6, pp. 475-477, 2016.[526] C. Guo, H. Wang, T. Jian, et al., “Radar target recognition based onfeature pyramid fusion lightweight CNN,”
IEEE Access , vol. 7, pp.51140-51149, 2019. Doi: 10.1109/ACCESS.2019.2909348.[527] D. Zhou, ”Radar target HRRP recognition based on reconstructive anddiscriminative dictionary learning,”
Signal Processing , pp. 52-64, 2016.[528] B. Xu, B. Chen, J. Wan, et al., “Target-aware recurrent attentionalnetwork for radar HRRP target recognition,”
Signal Processing , pp.268-280, 2019.[529] P. Mian, J. Jie, L. Zhu, et al., “Radar HRRP recognition basedon discriminant deep autoencoders with small training data size,” lectronics Letters , vol. 52, no. 20, pp. 1725-1727, 2016.[530] W. Xiong, G. Zhang, S. Liu, et al., “Multiscale kernel sparse coding-based classifier for HRRP radar target recognition,”
IET Radar Sonarand Navigation , vol. 10, no. 9, pp. 1594-1602, 2016.[531] Y. Guo, H. Xiao, Q. Fu, et al., “Least square support vector datadescription for HRRP-based radar target recognition,”
Applied Intel-ligence , vol. 46, no. 2, pp. 365-372, 2017.[532] B. Feng, B. Chen, H. Liu, et al., “Radar HRRP target recognition withdeep networks,”
Pattern Recognition , pp. 379-393, 2017.[533] Y. Kim, H. Ling, “ Human activity classification based on micro-doppler signatures using a support vector machine,”
IEEE Transactionson Geoscience and Remote Sensing , vol. 47, no. 5, pp. 1328-1337,2009.[534] Y. Kim, H. Ling, “ Human activity classification based on micro-Doppler signatures using an artificial neural network,” in
Proc. IEEEAntennas and Propagation Society International Symposium , 2008, pp.1-4.[535] Y. Kim, T. Moon, “Human detection and activity classification basedon micro-doppler signatures using deep convolutional neural networks,”
IEEE Geoscience and Remote Sensing Letters , vol. 13, no. 1, pp. 8-12,2016.[536] F. Fioranelli, M. Ritchie, H. D. Griffiths, et al., “Classification of un-armed/armed personnel using the NetRAD multistatic radar for micro-doppler and singular value decomposition features,”
IEEE Geoscienceand Remote Sensing Letters , vol. 12, no. 9, pp. 1933-1937, 2015.[537] Seyfioglu, M. Gurbuz, S.i Ozbayoglu, et al., “Deep learning of micro-Doppler features for aided and unaided gait recognition,” pp. 1125-1130, 2017.[538] F. Fioranelli, M. Ritchie, S. Z. Gurbuz, et al., “Feature diversity foroptimized human micro-doppler classification using multistatic Radar,”
IEEE Transactions on Aerospace and Electronic Systems , vol. 53, no.2, pp. 640-654, 2017.[539] J. Zhang, J. Tao, Z. Shi, et al., “Doppler-radar based hand gesturerecognition system using convolutional neural networks,” in
Proc.International Conference on Communications , 2017, pp. 1096-1113.[540] J. Ren, X. Jiang, “ Regularized 2-D complex-log spectral analysisand subspace reliability analysis of micro-doppler signature for UAVdetection,”
Pattern Recognition , pp. 225-237, 2017.[541] S. Abdulatif, F. Aziz, K. Armanious, “A study of human body charac-teristics effect on micro-doppler-based person identification using deeplearning,”
ArXiv : 1811.07173, 2019. [542] M. S. Seyfioglu, B. Erol, S. Z. Gurbuz, et al., “DNN transfer learningfrom diversified micro-doppler for motion classification,”
IEEE Trans-actions on Aerospace and Electronic Systems , vol. 55, no. 5, pp. 2164-2180, 2019.[543] M. S. Seyfioglu, A. M. Ozbayoglu, S. Z. Gurbuz, et al., “Deep con-volutional autoencoder for radar-based classification of similar aidedand unaided human activities,”
IEEE Transactions on Aerospace andElectronic Systems , vol. 54, no. 4, pp. 1709-1723, 2018.[544] M. S. Seyfiofilu, S. Z. Grbz, “Deep neural network initializationmethods for micro-doppler classification with low training samplesupport,”
IEEE Geoscience and Remote Sensing Letters , vol. 14, no.12, pp. 2462-2466, Dec. 2017.[545] K. N. Parashar, M. C. Oveneke, M. Rykunov, et al., “Micro-dopplerfeature extraction using convolutional auto-encoders for low latencytarget classification,” in
Proc. 2017 IEEE Radar Conference (Radar-Conf) , Seattle, WA, 2017, pp. 1739-1744.[546] A. W. Miller, C. Clemente, A. Robinson, et al., “Micro-doppler basedtarget classification using multi-feature integration,” in
Proc. 2013 IETIntelligent Signal Processing Conference (ISP 2013) , London, 2013,pp. 1-6.[547] G. Klarenbeek, R. I. A. Harmanny, L. Cifola, “Multi-target humangait classification using LSTM recurrent neural networks applied tomicro-Doppler,” in
Proc. 2017 European Radar Conference (EURAD) ,Nuremberg, 2017, pp. 167-170.[548] B. K. Kim, H. Kang, S. Park, “Drone classification using convolutionalneural networks with merged doppler images,”
IEEE Geoscience andRemote Sensing Letters , vol. 14, no. 1, pp. 38-42, Jan. 2017.[549] D. Tahmoush, “Review of micro-Doppler signatures,”
IET RadarSonar and Navigation , vol. 9, no. 9, pp. 1140-1146, 2015.[550] V. C. Chen, “Doppler signatures of radar backscattering from objectswith micro-motions,”
IET Signal Processing , vol. 2, no. 3, pp. 291-300,2008.[551] J. S. Patel, F. Fioranelli, D. Anderson, et al., “Review of radarclassification and RCS characterisation techniques for small UAVs ordrones,”
IET Radar Sonar and Navigation , vol. 12, no. 9, pp. 911-919,2018.[552] M. Ritchie, F. Fioranelli, H. Borrion, et al., “Multistatic micro-dopplerradar feature extraction for classification of unloaded/loaded micro-drones,”
IET Radar Sonar and Navigation , vol. 11, no. 1, pp. 116-124,2017.[553] S. Rahman, D. A. Robertson, ”Radar micro-Doppler signatures ofdrones and birds at K-band and W-band,”
Scientific Reports , vol. 8,no. 1, 2018.[554] X. Shi, F. Zhou, X. Bai, et al., “Deceptive jamming for tracked vehiclesbased on micro-Doppler signatures,”
IET Radar Sonar and Navigation ,vol. 12, no. 8, pp. 844-852, 2018.[555] T. Thayaparan, S. Abrol, E. Riseborough, et al., “Analysis of radarmicro-Doppler signatures from experimental helicopter and humandata,”
IET Radar Sonar and Navigation , vol. 1, no. 4, pp. 289-299,2007.[556] B. Jokanovic, M. Amin, “Fall detection using deep learning in range-doppler radars,”
IEEE Transactions on Aerospace and ElectronicSystems , vol. 54, no. 1, pp. 180-189, Feb. 2018.[557] S. Wang, J. Song, J. Lien, et al., “Interacting with soli: Exploring fine-grained dynamic gesture recognition in the radio-frequency spectrum,”in
Proc. User Interface Software and Technology , 2016, pp. 851-860.[558] Y. Li, Z. Peng, R. Pal, et al., “Potential active shooter detection based onradar micro-doppler and range-doppler analysis using artificial neuralnetwork,”
IEEE Sensors Journal , vol. 19, no. 3, pp. 1052-1063, 2019.[559] R. Zhang, S. Cao, “Support vector machines for classification ofautomotive radar interference,” in
Proc. 2018 IEEE Radar Confer-ence (RadarConf18) , Oklahoma City, OK, 2018, pp. 0366-0371. Doi:10.1109/RADAR.2018.8378586.[560] Ferre R M, La Fuente A D, Lohan E S, et al., “Jammer classificationin GNSS bands via machine learning algorithms,”
Sensors , vol. 19, no.22, 2019.[561] Z. Hao, W. Yu, W. Chen, “Recognition method of dense false targetsjamming based on time-frequency atomic decomposition,”
The Journalof Engineering , vol. 2019, no. 20, pp. 6354-6358, Oct. 2019. Doi:10.1049/joe.2019.0147.[562] A. Mendoza, A. Soto, B. C. Flores, “Classification of radar jam-mer FM signals using a neural network,” in
Pro. of SPIE , 2017.https://doi.org/10.1117/12.2262059[563] Y. Yang, L. Zhu, “An efficient way for satellite interference signalrecognition via incremental learning,” in
Proc. International Sympo-sium on Networks Computers and Communications , 2019, pp. 1-5.
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 48 [564] M. Kong, J. Liu, Z. Zhang, et al., “Radio ground-to-air interferencesignals recognition based on support vector machine,” in
Proc. Inter-national Conference on Digital Signal Processing , 2015, pp. 987-990.[565] F. Ruo-Ran, “Compound jamming signal recognition based on neuralnetworks,” in
Proc. 2016 Sixth International Conference on Instru-mentation and Measurement, Computer, Communication and Control(IMCCC) , Harbin, 2016, pp. 737-740. Doi: 10.1109/IMCCC.2016.163.[566] Z. Wu, Y. Zhao, Z. Yin, et al., “Jamming signals classification usingconvolutional neural network,” in
Proc. 2017 IEEE International Sym-posium on Signal Processing and Information Technology (ISSPIT) ,Bilbao, 2017, pp. 062-067. Doi: 10.1109/ISSPIT.2017.8388320.[567] Q. Liu, W. Zhang, “Deep learning and recognition of radar jammingbased on CNN,” in
Proc. 2019 12th International Symposium onComputational Intelligence and Design (ISCID) , Hangzhou, China,2019, pp. 208-212. Doi: 10.1109/ISCID.2019.00054.[568] Y. Cai, K. Shi, F. Song, et al., “Jamming pattern recognition us-ing spectrum waterfall: A deep learning method,” in
Proc. 2019IEEE 5th International Conference on Computer and Commu-nications (ICCC) , Chengdu, China, 2019, pp. 2113-2117. Doi:10.1109/ICCC47050.2019.9064207.[569] ”Research on electronic jamming identification based on CNN,” in
Proc. 2019 IEEE interference on signal, data and signal processing ,Chongqing, 2019.[570] L. Tingpeng, W. Manxi, P. Danhua, et al., “Identification of jammingfactors in electronic information system based on deep learning,”in
Proc. 2018 IEEE 18th International Conference on Communi-cation Technology (ICCT) , Chongqing, 2018, pp. 1426-1430, doi:10.1109/ICCT.2018.8600065[571] Y. Wang, B. Sun, N. Wang, “Recognition of radar active-jammingthrough convolutional neural networks,”
The Journal of Engi-neering , vol. 2019, no. 21, pp. 7695-7697, Nov. 2019. Doi:10.1049/joe.2019.0659.[572] X. Tang, X. Zhang, J. Shi, et al., “SAR deception jamming targetrecognition based on the shadow feature,” in
Proc. 2017 25th EuropeanSignal Processing Conference (EUSIPCO) , Kos, 2017, pp. 2491-2495.Doi: 10.23919/EUSIPCO.2017.8081659.[573] M. Zhu, Y. Li, Z. Pan, et al., “Automatic modulation recognition ofcompound signals using a deep multi-label classifier: A case study withradar jamming signals,”
Signal Processing , 2020.[574] Lee G, Jo J, Park C H, et al, “Jamming prediction for radar signals usingmachine learning methods,”
Security and Communication Networks ,pp. 1-9, 2020.[575] Q. Wu, Z. Sun, X. Zhou, “Interference detection and recognition basedon signal reconstruction using recurrent neural network,” in
Proc. 2019IEEE Globecom Workshops (GC Wkshps) , Waikoloa, HI, USA, 2019,pp. 1-6. Doi: 10.1109/GCWkshps45667.2019.9024542.[576] L. Kang, J. Bo, L. Hongwei, et al., “Reinforcement learning basedanti-jamming frequency hopping strategies design for cognitive radar,”in
Proc. International Conference on Signal Processing , 2018.[577] C. Han, L. Huo, X. Tong, et al., “Spatial anti-jamming scheme forinternet of satellites based on the deep reinforcement learning andStackelberg game,”
IEEE Transactions on Vehicular Technology , vol.69, no. 5, pp. 5331-5342, May 2020. Doi: 10.1109/TVT.2020.2982672.[578] P. Liu, Y. Liu, T. Huang, et al., “Decentralized automotive radarspectrum allocation to avoid mutual interference using reinforcementlearning,”
ArXiv : Signal Processing, 2020.[579] J. Mun, H. Kim, J. Lee, “A deep learning approach for automotiveradar interference mitigation,” in
Proc. 2018 IEEE 88th VehicularTechnology Conference (VTC-Fall) , Chicago, IL, USA, 2018, pp. 1-5. Doi: 10.1109/VTCFall.2018.8690848.[580] L. Wang, S. Fortunati, M. Greco, et al., “Reinforcement learning-basedwaveform optimization for MIMO multi-target detection,” in
Proc.Asilomar Conference on Signals, Systems and Computers , 2018, pp.1329-1333.[581] J. Wei, M. H. Alexander, S. Osvaldo, “End-to-end learning of waveformgeneration and detection for radar system,”
ArXiv : 1912.00802v1, 2019.[582] O. Naparstek, K. Cohen, “Deep multi-user reinforcement learning fordistributed dynamic spectrum access,”
IEEE Transactions on WirelessCommunications , vol. 18, no. 1, pp. 310-323, 2019.[583] L. R. Faganello, R. Kunst, C. B. Both, et al., “Improving reinforcementlearning algorithms for dynamic spectrum allocation in cognitivesensor networks,” in
Proc. Wireless Communications and NetworkingConference , 2013, pp. 35-40.[584] S. You, M. Diao, L. Gao, et al., “Deep reinforcement learning for targetsearching in cognitive electronic warfare,”
IEEE Access , pp. 37432-37447, 2019. [585] J. G. Metcalf, S. D. Blunt, B. Himed, et al., “A machine learningapproach to cognitive radar detection,” in
Proc. IEEE radar conference ,2015, pp. 1405-1411.[586] A. M. Elbir, K. V. Mishra, Y. C. Eldar, et al., “Cognitive radar antennaselection via deep learning,”
IET Radar Sonar and Navigation , vol.13, no. 6, pp. 871-880, 2019.[587] Z. Liu, D. K. Ho, X. Xu, et al., “Moving target indication using deepconvolutional neural network,”
IEEE Access , pp. 65651-65660, 2018.[588] X. B. Liu, Y. A. Xu, H. Ding, et al., “High-dimensional featureextraction of sea clutter and target signal for intelligent maritimemonitoring network,”
Computer Communications , vol. 147, pp. 76-84,2019.[589] M. P. Jaraboamores, M. Rosazurera, R. Gilpita, et al., “Study oftwo error functions to approximate the NeymanCPearson detectorusing supervised learning machines,”
IEEE Transactions on SignalProcessing , vol. 57, no. 11, pp. 4175-4181, 2009.[590] Y. Q. Wang, Q. M. Yao. “Few-shot Learning: A Survey,”
ArXiv :1904.05046v1, 2020.[591] S. Ravi, H. Larochelle, “Optimization as a model for few-shot learn-ing,” in
Proc. International Conference on Learning Representations(ICLR) , 2017.[592] R. Socher, M. Ganjoo, C. D. Manning, et al., “Zero-shot learningthrough cross-modal transfer,” in
Proc. Neural Information ProcessingSystems , 2013, PP. 935-943.[593] Z. M. Ding., H. D. Zhao, Y. Fu. “Zero-Shot Learning,” 2019.doi:10.1007/978-3-030-00734-8-6.[594] T. Elsken, J. H. Metzen, F. Hutter, et al., “Neural architecture search:A survey,”
Journal of Machine Learning Research , vol. 20, no. 55, pp.1-21, 2019.[595] Y. Cheng, D. Wang, P. Zhou, et al., “A survey of model compressionand acceleration for deep neural networks,”
ArXiv: Learning , 2017.[596] D. Castelvecchi, “Can we open the black box of AI,”
Nature , vol. 538,no. 7623, pp. 20-23, 2016.[597] A. Adadi, M. Berrada, “Peeking inside the black-box: A survey onexplainable artificial intelligence (XAI),”
IEEE Access , pp. 52138-52160, 2018.[598] E. Tjoa, C. Guan, “A survey on explainable artificial intelligence (XAI):Towards medical XAI,”
ArXiv: Learning , 2019.[599] M. T. Ribeiro, S. Singh, C. Guestrin, et al., “Why should I trustyou?: Explaining the predictions of any classifier,” in
Proc. KnowledgeDiscovery and Data Mining , 2016, pp. 1135-1144.[600] S. Tan, R. Caruana, G. Hooker, et al., “Learning global additiveexplanations for neural nets using model distillation,”
ArXiv: MachineLearning , 2018.[601] T. Wu, W. Sun, X. Li, et al., “Towards interpretable R-CNN by unfold-ing latent structures,”
ArXiv: Computer Vision and Pattern Recognition ,2017.[602] M. D. Zeiler, R. Fergus,” “Visualizing and understanding convolutionalnetworks,” in
Proc. European Conference on Computer Vision , 2014,pp. 818-833.[603] J. Yosinski, J. Clune, A. Nguyen, et al., “Understanding neural net-works through deep visualization,”
ArXiv: Computer Vision and PatternRecognition , 2015.[604] L. M. Zintgraf, T. S. Cohen, T. Adel, et al., “Visualizing deepneural network decisions: Prediction difference analysis,” in
Proc.International Conference on Learning Representations (ICLR) , 2017.[605] P. Kindermans, K. T. Schutt, M. Alber, et al., “Learning how toexplain neural networks: PatternNet and PatternAttribution,” in
Proc.International Conference on Learning Representations (ICLR) , 2018.[606] M. Ancona, E. Ceolini, C. Oztireli, et al., “Towards better understand-ing of gradient-based attribution methods for Deep Neural Networks,”in
Proc. International Conference on Learning Representations (ICLR) ,2018, pp. 0-0.[607] B. Letham, C. Rudin, T. H. Mccormick, et al., “Interpretable classifiersusing rules and Bayesian analysis: Building a better stroke predictionmodel,”
The Annals of Applied Statistics , vol. 9, no. 3, pp. 1350-1371,2015.[608] V. Joel, S. Agus, et al., “Explainable neural networks based on additiveindex models,”
ArXiv : 1806.01933, 2018.[609] A. M. David, J. Tommi, “Towards robust interpretability with self-explaining neural networks,” in
Proc. Neural Information ProcessingSystems (NIPS) , 2018, pp. 7786-7795.[610] A. S. Maruan, D. Avinava, P. X. Eric, “Contextual explanation net-works,”
ArXiv : 1705.10301, 2018.[611] Q. S. Zhang, Y. N. Wu, S. C. Zhu, “Interpretable convolutional neuralnetworks,”
ArXiv : 1710.00935, 2018.
OURNAL OF L A TEX CLASS FILES, VOL. X, NO. X, SEP 2020 49 [612] T. D. Kulkarni, W. F. Whitney, P. Kohli, et al., “Deep convolutionalinverse graphics network,” in
Proc. Neural Information ProcessingSystems (NIPS) , 2015, pp. 2539-2547.[613] K. Hyunjik, M. Andry, “Disentangling by factorising,”
ArXiv :1802.05983v3, 2019.[614] M. David, T. Philip, et al., “Transparency by design: Closing the gapbetween performance and interpretability in visual reasoning,”
ArXiv :1803.05268, 2018.[615] Xiaoran X., Songpeng Z., and et al. Modeling Attention Flow onGraphs. arXiv:1811.00497, 2018.[616] N. Tishby, N. Zaslavsky, “Deep learning and the information bottleneckprinciple,” in
Proc. Information Theory Workshop , 2015, pp. 1-5.[617] R. Shwartzziv, N. Tishby, “Opening the black box of deep neuralnetworks via information,”
ArXiv: Learning , 2017.[618] Q. Li, B. Wang, M. Melucci, et al., “CNM: An interpretable complex-valued network for matching,”
ArXiv: Computation and Language ,2019.[619] Y. P. Lu, Z. H. Li, et al., “Understanding and improving transformerfrom a multi-particle dynamic system point of view,”
Arxiv : 190602762, 2019.[620] P. Vincke, “Analysis of multicriteria decision aid in europe,”
EuropeanJournal of Operational Research , vol. 25, no. 2, pp. 160-168, 1986.[621] M.Z. Guo, Q.P. Zhang, X.W Liao, et al., “An interpretable machinelearning framework for modelling human decision behavior,”
ArXiv :1906.01233, 2019.[622] B. M. Lake, R. Salakhutdinov, J. B. Tenenbaum, et al., “Human-levelconcept learning through probabilistic program induction,”
Science ,vol. 350, no. 6266, pp. 1332-1338, 2015.[623] S. Murray, N. Kyriacos, C. Antonia, et al., “An explicitly relationalneural network architecture,”
ArXiv : 1905.10307v1, 2019.[624] J. Y. Mao, C. Gan, K. Pushmeet, et al., “The neuro-symbolic conceptlearner: Interpreting scenes, words, and sentences from natural super-vision,” in
Proc. International Conference on Learning Representation(ICLR) , 2019.[625] D. Goldfarb, S. C. Evans, “Causal inference via conditional Kol-mogorov complexity using MDL Binning,”
ArXiv: Learning , 2019.[626] Y. Bengio, “Deep learning of representations: Looking forward,” in
Proc. International Conference on Statistical Language and SpeechProcessing ,Springer, 2013, pp. 1-37.[627] M. Garnelo, M. Shanahan, “Reconciling deep learning with symbolicartificial intelligence: representing objects and relations,” in
Proc.Current Opinion in Behavioral Sciences 29 , 2019, pp. 17C23.[628] Y. Bengio, “The consciousness prior,”
ArXiv : 1709.08568, 2017.[629] G.N.Rao, C.V.S.Sastry, N.Divakar, “Trends in electronic warfare,”
IETE Tech. Rev. , vol. 20, no. 2, pp. 139C150, Feb. 2003.[630] S. Haykin, “Cognitive radar: a way of the future,”
IEEE SignalProcessing Magazine , vol. 23, no. 1, pp. 30-40, 2006.[631] S. Haykin, Y. Xue, P. Setoodeh, “Cognitive radar: Step towardbridging the gap between neuroscience and engineering,” in
Proc.of the IEEE , vol. 100, no. 11, pp. 3102-3130, Nov. 2012. Doi:10.1109/JPROC.2012.2203089.[632] J. R. Guerci, “Cognitive radar: A knowledge-aided fully adaptiveapproach,” in
Proc. 2010 IEEE Radar Conference , Washington, DC,2010, pp. 1365-1370. Doi: 10.1109/RADAR.2010.5494403.[633] W. Huleihel, J. Tabrikian, R. Shavit, et al., “Optimal adaptive waveformdesign for cognitive MIMO Radar,”
IEEE Transactions on SignalProcessing , vol. 61, no. 20, pp. 5075-5089, 2013.[634] A. Lodi, G. Zarpellon, “On learning and branching: A survey,”
Top ,vol. 25, no. 2, pp. 207-236, 2017.[635] N. Balcan, T. Dick, T. Sandholm, et al., “Learning to Branch,” in
Proc.International Conference on Machine Learning (ICML) , 2018, pp. 344-353.[636] Y. Tang, S. Agrawal, Y. Faenza, et al., “Reinforcement learning forinteger programming: Learning to cut,”
ArXiv: Learning , 2019.[637] Y. Bengio, A. Lodi, A. Prouvost, et al., “Machine learning forcombinatorial optimization: A methodological tour d’Horizon,”
ArXiv:Learning , 2018.[638] H. J. Dai, B. K. Elias, Z. Yuyu, et al., “Learning combinatorialoptimization algorithms over graphs,”
ArXiv : 1704.01665v4, 2018.[639] M. Nazari, A. Oroojlooy, L. V. Snyder, et al., “Reinforcement learningfor solving the vehicle routing problem,” in
Proc. Neural InformationProcessing Systems (NIPS) , 2018, PP. 9861-9871.[640] O. Vinyals, M. Fortunato, N. Jaitly, et al., “Pointer networks,” in