Daniël Reichman

On Choosing Training and Testing Data for Supervised Algorithms in Ground-Penetrating Radar Data for Buried Threat Detection

Ground-penetrating radar (GPR) is one of the most popular and successful sensing modalities that have been investigated for landmine and subsurface threat detection. Many of the detection algorithms applied to this task are supervised and therefore require labeled examples of threat and nonthreat data for training. Training data most often consist of 2-D images (or patches) of GPR data, from which features are extracted and provided to the classifier during training and testing. Identifying desirable training and testing locations to extract patches, which we term “keypoints,” is well established in the literature. In contrast, however, a large variety of strategies have been proposed regarding keypoint utilization (e.g., how many of the identified keypoints should be used at threat, or nonthreat, locations). Given a variety of keypoint utilization strategies that are available, it is very unclear: 1) which strategies are best or 2) whether the choice of strategy has a large impact on classifier performance. We address these questions by presenting a taxonomy of existing utilization strategies and then evaluating their effectiveness on a large data set using many different classifiers and features. We analyze the results and propose a new strategy, called PatchSelect, which outperforms other strategies across all experiments.

Target signature localization in GPR data by jointly estimating and matching templates

Buried threat detection algorithms in Ground Penetrating Radar (GPR) measurements often utilize a statistical classifier to model target responses. There are many different target types with distinct responses and all are buried in a wide range of conditions that distort the target signature. Robust performance of this classifier requires it to learn the distinct responses of target types while accounting for the variability due to the physics of the emplacement. In this work, a method to reduce certain sources of excess variation is presented that enables a linear classifier to learn distinct templates for each target type’s response despite the operational variability. The different target subpopulations are represented by a Gaussian Mixture Model (GMM). Training the GMM requires jointly extracting the patches around target responses as well as learning the statistical parameters as neither are known a priori. The GMM parameters and the choice of patches are determined by variational Bayesian methods. The proposed method allows for patches to be extracted from a larger data-block that only contain the target response. The patches extracted from this method improve the ROC for distinguishing targets from background clutter compared to the patches extracted using other patch extraction methods aiming to reduce the operational variability.

Target localization and signature extraction in GPR data using expectation-maximization and principal component analysis

Ground Penetrating Radar (GPR) is a very promising technology for subsurface threat detection. A successful algorithm employing GPR should achieve high detection rates at a low false-alarm rate and do so at operationally relevant speeds. GPRs measure reflections at dielectric boundaries that occur at the interfaces between different materials. These boundaries may occur at any depth, within the sensors range, and furthermore, the dielectric changes could be such that they induce a 180 degree phase shift in the received signal relative to the emitted GPR pulse. As a result of these time-of-arrival and phase variations, extracting robust features from target responses in GPR is not straightforward. In this work, a method to mitigate polarity and alignment variations based on an expectation-maximization (EM) principal-component analysis (PCA) approach is proposed. This work demonstrates how model-based target alignment can significantly improve detection performance. Performance is measured according to the improvement in the receiver operating characteristic (ROC) curve for classification before and after the data is properly aligned and phase-corrected.

Some good practices for applying convolutional neural networks to buried threat detection in Ground Penetrating Radar

Ground Penetrating Radar (GPR) is a remote sensing modality that has been researched extensively for buried threat detection. For this purpose, algorithms can be developed to automatically determine the presence of such threats. To train such algorithms, small 2-dimensional images can be extracted from the larger image, or volume, of GPR data. One thread of research in the buried threat detection literature is to use visual descriptors from the computer vision literature. One recent, very successful approach in that field is the use of deep convolutional neural networks (CNNs). Applying CNNs requires a large number of design choices which complicate their use. In this work, we investigate their application to GPR data and adapt several recent advances from the CNN literature to improve detection performance on GPR data. In particular, we investigate the initialization step of pretraining and propose a dataset augmentation protocol. The efficacy of these approaches are evaluated on several architectures with a relatively similar number of network parameters to learn. The results indicate that both pretraining and dataset augmentation help achieve higher detection performance.

Discriminative dictionary learning to learn effective features for detecting buried threats in ground penetrating radar data

The ground penetrating radar (GPR) is a popular remote sensing modality for buried threat detection. In this work we focus on the development of supervised machine learning algorithms that automatically identify buried threats in GPR data. An important step in many of these algorithms is feature extraction, where statistics or other measures are computed from the raw GPR data, and then provided to the machine learning algorithms for classification. It is well known that an effective feature can lead to major performance improvements and, as a result, a variety of features have been proposed in the literature. Most of these features have been handcrafted, or designed through trial and error experimentation. Dictionary learning is a class of algorithms that attempt to automatically learn effective features directly from the data (e.g., raw GPR data), with little or no supervision. Dictionary learning methods have yielded state-of-theart performance on many problems, including image recognition, and in this work we adapt them to GPR data in order to learn effective features for buried threat classification. We employ the LC-KSVD algorithm, which is a discriminative dictionary learning approach, as opposed to a purely reconstructive one like the popular K-SVD algorithm. We use a large collection of GPR data to show that LC-KSVD outperforms two other approaches: the popular Histogram of oriented gradient (HOG) with a linear classifier, and HOG with a nonlinear classifier (the Random Forest).

Learning improved pooling regions for the Histogram of Oriented Gradient (HOG) feature for buried threat detection in ground penetrating radar

In recent years, the Ground Penetrating Radar (GPR) has successfully been applied to the problem of buried threat detection (BTD). A large body of research has focused on using computerized algorithms to automatically discriminate between buried threats and subsurface clutter in GPR data. For this purpose, the GPR data is frequently treated as an image of the subsurface, within which the reflections associated with targets often appear with a characteristic shape. In recent years, shape descriptors from the natural image processing literature have been applied to buried threat detection, and the histogram of oriented gradient (HOG) feature has achieved state-of-the-art performance. HOG consists of computing histograms of the image gradients in disjoint square regions, which we call pooling regions, across the GPR images. In this work we create a large body of potential pooling regions and use the group LASSO (GLASSO) to choose a subset of the pooling regions that are most appropriate for BTD on GPR data. We examined this approach on a large collection of GPR data using lane-based cross-validation, and the results indicate that GLASSO can select a subset of pooling regions that lead to superior performance to the original HOG feature, while simultaneously also reducing the total number of features needed. The selected pooling regions also provide insight about the regions in GPR images that are most important for discriminating threat and nonthreat data.

Algorithm development for deeply buried threat detection in GPR data

Ground penetrating radar (GPR) is a popular remote sensing modality for buried threat detection. Many algorithms have been developed to detect buried threats using GPR data. One on-going challenge with GPR is the detection of very deeply buried targets. In this work a detection approach is proposed that improves the detection of very deeply buried targets, and interestingly, shallow targets as well. First, it is shown that the signal of a target (the target “signature”) is well localized in time, and well correlated with the target’s burial depth. This motivates the proposed approach, where GPR data is split into two disjoint subsets: an early and late portion corresponding to the time at which shallow and deep target signatures appear, respectively. Experiments are conducted on real GPR data using the previously published histogram of oriented gradients (HOG) prescreener: a fast supervised processing method operated on HOG features. The results show substantial improvements in detection of very deeply buried targets (4.1% to 17.2%) and in overall detection performance (81.1% to 83.9%). Further, it is shown that the performance of the proposed approach is relatively insensitive to the time at which the data is split. These results suggest that other detection methods may benefit from depth-based processing as well.

If training data appears to be mislabeled, should we relabel it? Improving supervised learning algorithms for threat detection in ground penetrating radar data

This work focuses on the development of automatic buried threat detection (BTD) algorithms using ground penetrating radar (GPR) data. Buried threats tend to exhibit unique characteristics in GPR imagery, such as high energy hyperbolic shapes, which can be leveraged for detection. Many recent BTD algorithms are supervised, and therefore they require training with exemplars of GPR data collected over non-threat locations and threat locations, respectively. Frequently, data from non-threat GPR examples will exhibit high energy hyperbolic patterns, similar to those observed from a buried threat. Is it still useful therefore, to include such examples during algorithm training, and encourage an algorithm to label such data as a non-threat? Similarly, some true buried threat examples exhibit very little distinctive threat-like patterns. We investigate whether it is beneficial to treat such GPR data examples as mislabeled, and either (i) relabel them, or (ii) remove them from training. We study this problem using two algorithms to automatically identify mislabeled examples, if they are present, and examine the impact of removing or relabeling them for training. We conduct these experiments on a large collection of GPR data with several state-of-the-art GPR-based BTD algorithms.

How much shape information is enough, or too much? Designing imaging descriptors for threat detection in ground penetrating radar data

In this work, we consider the development of algorithms for automated buried threat detection (BTD) using Ground Penetrating Radar (GPR) data. When viewed in GPR imagery, buried threats often exhibit hyperbolic shapes, and this characteristic shape can be leveraged for buried threat detection. Consequentially, many modern detectors initiate processing the received data by extracting visual descriptors of the GPR data (i.e., features). Ideally, these descriptors succinctly encode all decision-relevant information, such as shape, while suppressing spurious data content (e.g., random noise). Some notable examples of successful descriptors include the histogram of oriented gradient (HOG), and the edge histogram descriptor (EHD). A key difference between many descriptors is the precision with which shape information is encoded. For example, HOG encodes shape variations over both space and time (high precision); while EHD primarily encodes shape variations only over space (lower precision). In this work, we conduct experiments on a large GPR dataset that suggest EHD-like descriptors outperform HOG-like descriptors, as well as exhibiting several other practical advantages. These results suggest that higher resolution shape information (particularly shape variations over time) is not beneficial for buried threat detection. Subsequent analysis also indicates that the performance advantage of EHD is most pronounced among difficult buried threats, which also exhibit more irregular shape patterns.

Improving the histogram of oriented gradient feature for threat detection in ground penetrating radar by implementing it as a trainable convolutional neural network

A large number of algorithms have been proposed for automatic buried threat detection (BTD) in ground penetrating radar (GPR) data. Convolutional neural networks (CNNs) have recently achieved groundbreaking results on many recognition tasks. This success is due, in part, to their ability to automatically infer effective data representations (i.e., features) using training data. This capability however results in a high capacity model (i.e., many free parameters) that is difficult to train, and more prone to overfitting, than models employing hand-crafted feature designs. This drawback is pronounced when training data is relatively scarce, as is the case with GPR BTD. In this work we propose to combine the relative advantages of hand-crafted features, and CNNs, by constructing CNN architectures that closely emulate successful hand-crafted feature designs for GPR BTD. This makes it possible to apply supervised training to traditional hand-crafted features, allowing them to adapt to the unique characteristics of the GPR BTD problem. Simultaneously, this approach yields a much lower capacity CNN model that incorporates substantial prior research knowledge, making the model much easier to train. We demonstrate the feasibility and effectiveness of this approach by designing a “neural” implementation of the popular histogram of oriented gradient (HOG) feature. The resulting neural HOG (NHOG) implementation is much smaller and easier to train than standard CNN architectures, and achieves superior detection performance compared to the un-trained HOG feature. In theory, neural implementations can be developed for many existing successful GPR BTD algorithms, potentially yielding similar benefits.