Erzhu Li

Hyperspectral Image Classification Using Band Selection and Morphological Profiles

In this paper, we propose a simple unsupervised framework to effectively select and combine spectral information and spatial features for Support Vector Machine (SVM)-based classification when spatial features are the widely used morphological profiles (MPs). To overcome the difficulty of high dimensionality of resulting features, it is a common practice that MPs are extracted from principal components (PCs). In this paper, we investigate another technique on spectral feature selection, which is unsupervised band selection (BS). We find out that using selected bands as spectral features can improve classification performance because they contain more critical characteristics for classification; in particular, using the selected bands, combined with the MPs extracted from PCs, can yield the highest accuracy, due to the fact that major PCs contain less noise for extracting more reliable MPs. The overall unsupervised nature of feature selection provides the flexibility of implementation. We believe that such finding is instructive to feature selection and extraction for spectral/spatial-based hyperspectral image classification.

Integrating Multilayer Features of Convolutional Neural Networks for Remote Sensing Scene Classification

Scene classification from remote sensing images provides new possibilities for potential application of high spatial resolution imagery. How to efficiently implement scene recognition from high spatial resolution imagery remains a significant challenge in the remote sensing domain. Recently, convolutional neural networks (CNN) have attracted tremendous attention because of their excellent performance in different fields. However, most works focus on fully training a new deep CNN model for the target problems without considering the limited data and time-consuming issues. To alleviate the aforementioned drawbacks, some works have attempted to use the pretrained CNN models as feature extractors to build a feature representation of scene images for classification and achieved successful applications including remote sensing scene classification. However, existing works pay little attention to exploring the benefits of multilayer features for improving the scene classification in different aspects. As a matter of fact, the information hidden in different layers has great potential for improving feature discrimination capacity. Therefore, this paper presents a fusion strategy for integrating multilayer features of a pretrained CNN model for scene classification. Specifically, the pretrained CNN model is used as a feature extractor to extract deep features of different convolutional and fully connected layers; then, a multiscale improved Fisher kernel coding method is proposed to build a mid-level feature representation of convolutional deep features. Finally, the mid-level features extracted from convolutional layers and the features of fully connected layers are fused by a principal component analysis/spectral regression kernel discriminant analysis method for classification. For validation and comparison purposes, the proposed approach is evaluated via experiments with two challenging high-resolution remote sensing data sets, and shows the competitive performance compared with fully trained CNN models, fine-tuning CNN models, and other related works.

Mid-Level Feature Representation via Sparse Autoencoder for Remotely Sensed Scene Classification

Feature representation is a classic problem in the machine learning community due to the fact that different representations can entangle and hide more or less the different explanatory factors of variation behind the raw data. Especially for scene classification, its performance generally depends on the discriminative power of feature representation. Recently, unsupervised feature learning attracts tremendous attention because of its ability to learn feature representation automatically. However, reliable performance of feature representations by unsupervised learning always requires a large number of features and complex framework of mid-level feature representation. To alleviate such drawbacks, this paper presents a new framework of mid-level feature representation, which does not need learn many convolutional features during the unsupervised feature learning process, and has few parameter settings. In detail, the unsupervised feature learning method, sparse autoencoder, is employed to learn relatively small number of convolutional features from input dataset, and then extended features are extracted from the learned features by a multiple normalized difference features extraction method to compose a derivative feature set. At mid-level feature representation stage, in order to avoid poor performance of standard pooling technology in solving problems brought by rotation and translation of scene images, global feature descriptors (histogram moments, mean, variance, standard deviation) are utilized to build mid-level feature representations of images. For validation and comparison purposes, the proposed approach is evaluated via experiments with two challenging high-resolution remote sensing datasets. The results demonstrate that the approach is effective, and shows strong performance for remotely sensed scene classification.

An automatic approach for urban land-cover classification from Landsat-8 OLI data

Due to the lack of clear shape, texture characteristics, and abundant spectral or spatial information of urban objects, traditional per-/sub-pixel analysis and interpretation for moderate-resolution-remote sensing data are always confused by such shortcomings as dependence on special skills, requirements to a priori knowledge and training samples, complex process, time-consuming and subjective operations, etc.. In order to alleviate such disadvantages, an automatic approach is proposed to classify vegetation, water, impervious surface areas (dark and bright), and bare land from the Operational Land Imager (OLI) sensor data of Landsat-8 in urban areas, which can be employed by common users to automatically obtain land-cover maps for urban applications. In detail, a preliminary classification result is achieved based on a new vegetation and water masking index (VWMI), the normalized difference vegetation index (NDVI), and a new normalized difference bare land index (NDBLI), which are acquired automatically from the remote-sensing images based on available knowledge of spectral properties. VWMI is designed to extract vegetation and water information together with a simpler threshold, while NDBLI is developed to identify dark impervious surfaces and bare land in this work. A modification strategy is further proposed to improve preliminary classification results by a linear model. For this purpose, a stable sample selection method based on the histogram is developed to select training samples from the preliminary classification result and to build a non-linear support vector machine (SVM) model to reclassify the classes. For validation and comparison purposes, the proposed approach is evaluated via experiments with real OLI data of two study areas, Nanjing and Ordos. The results demonstrate that the approach is effective in automatically obtaining urban land-cover classification maps from OLI data for thematic analysis.

Supervised and Semi-Supervised Multi-View Canonical Correlation Analysis Ensemble for Heterogeneous Domain Adaptation in Remote Sensing Image Classification

In this paper, we present the supervised multi-view canonical correlation analysis ensemble (SMVCCAE) and its semi-supervised version (SSMVCCAE), which are novel techniques designed to address heterogeneous domain adaptation problems, i.e., situations in which the data to be processed and recognized are collected from different heterogeneous domains. Specifically, the multi-view canonical correlation analysis scheme is utilized to extract multiple correlation subspaces that are useful for joint representations for data association across domains. This scheme makes homogeneous domain adaption algorithms suitable for heterogeneous domain adaptation problems. Additionally, inspired by fusion methods such as Ensemble Learning (EL), this work proposes a weighted voting scheme based on canonical correlation coefficients to combine classification results in multiple correlation subspaces. Finally, the semi-supervised MVCCAE extends the original procedure by incorporating multiple speed-up spectral regression kernel discriminant analysis (SRKDA). To validate the performances of the proposed supervised procedure, a single-view canonical analysis (SVCCA) with the same base classifier (Random Forests) is used. Similarly, to evaluate the performance of the semi-supervised approach, a comparison is made with other techniques such as Logistic label propagation (LLP) and the Laplacian support vector machine (LapSVM). All of the approaches are tested on two real hyperspectral images, which are considered the target domain, with a classifier trained from synthetic low-dimensional multispectral images, which are considered the original source domain. The experimental results confirm that multi-view canonical correlation can overcome the limitations of SVCCA. Both of the proposed procedures outperform the ones used in the comparison with respect to not only the classification accuracy but also the computational efficiency. Moreover, this research shows that canonical correlation weighted voting (CCWV) is a valid option with respect to other ensemble schemes and that because of their ability to balance diversity and accuracy, canonical views extracted using partially joint random view generation are more effective than those obtained by exploiting disjoint random view generation.

Assessing the impact of urbanization on net primary productivity using multi-scale remote sensing data: a case study of Xuzhou, China

An improved Carnegie Ames Stanford Approach (CASA) model based on two kinds of remote sensing (RS) data, Landsat Enhanced Thematic Mapper Plus (ETM +) and Moderate Resolution Imaging Spectro-radiometer (MODIS), and climate variables were applied to estimate the Net Primary Productivity (NPP) of Xuzhou in June of each year from 2001 to 2010. The NPP of the study area decreased as the spatial scale increased. The average NPP of terrestrial vegetation in Xuzhou showed a decreasing trend in recent years, likely due to changes in climate and environment. The study area was divided into four sub-regions, designated as highest, moderately high, moderately low, and lowest in NPP. The area designated as the lowest sub-region in NPP increased with expanding scale, indicating that the NPP distribution varied with different spatial scales. The NPP of different vegetation types was also significantly influenced by scale. In particular, the NPP of urban woodland produced lower estimates because of mixed pixels. Similar trends in NPP were observed with different RS data. In addition, expansion of residential areas and reduction of vegetated areas were the major reasons for NPP change. Land cover changes in urban areas reduced NPP, which could chiefly be attributed to human-induced disturbance.

Fuzzy multiclass active learning for hyperspectral image classification

The possibility theory, which is an extension of fuzzy sets and fuzzy logic, has shown considerable potential for solving active learning (AL) problems, particularly for multiclass scenarios’ classification. Hence, two recently proposed fuzzy multiclass AL algorithms (classification ambiguity (CA) and fuzzy C-order ambiguity (FCOA)) are investigated to properly generalise them for classifying hyperspectral images, and two improved versions of the CA and FCOA are proposed. In addition to comparing the performances of the original and improved algorithms, several other state-of-the-art AL methods are evaluated, such as breaking ties, margin sampling, and multi-class level uncertainty, with or without diversity criteria such as angle-based diversity (ABD), clustering-based diversity (CBD), and enhanced clustering-based diversity (ECBD). Tests on two benchmark hyperspectral images confirm that the proposed improved algorithms are superior to and more effective than the original ones.

Performance Evaluation of Downscaling Sentinel-2 Imagery for Land Use and Land Cover Classification by Spectral-Spatial Features

Land Use and Land Cover (LULC) classification is vital for environmental and ecological applications. Sentinel-2 is a new generation land monitoring satellite with the advantages of novel spectral capabilities, wide coverage and fine spatial and temporal resolutions. The effects of different spatial resolution unification schemes and methods on LULC classification have been scarcely investigated for Sentinel-2. This paper bridged this gap by comparing the differences between upscaling and downscaling as well as different downscaling algorithms from the point of view of LULC classification accuracy. The studied downscaling algorithms include nearest neighbor resampling and five popular pansharpening methods, namely, Gram-Schmidt (GS), nearest neighbor diffusion (NNDiffusion), PANSHARP algorithm proposed by Y. Zhang, wavelet transformation fusion (WTF) and high-pass filter fusion (HPF). Two spatial features, textural metrics derived from Grey-Level-Co-occurrence Matrix (GLCM) and extended attribute profiles (EAPs), are investigated to make up for the shortcoming of pixel-based spectral classification. Random forest (RF) is adopted as the classifier. The experiment was conducted in Xitiaoxi watershed, China. The results demonstrated that downscaling obviously outperforms upscaling in terms of classification accuracy. For downscaling, image sharpening has no obvious advantages than spatial interpolation. Different image sharpening algorithms have distinct effects. Two multiresolution analysis (MRA)-based methods, i.e., WTF and HFP, achieve the best performance. GS achieved a similar accuracy with NNDiffusion and PANSHARP. Compared to image sharpening, the introduction of spatial features, both GLCM and EAPs can greatly improve the classification accuracy for Sentinel-2 imagery. Their effects on overall accuracy are similar but differ significantly to specific classes. In general, using the spectral bands downscaled by nearest neighbor interpolation can meet the requirements of regional LULC applications, and the GLCM and EAPs spatial features can be used to obtain more precise classification maps.

Hyperspectral image classification using band selection and morphological profile

In this paper, we propose a new methodology to combine spectral information and spatial features for Support Vector Machine (SVM)-based classification. The novelty of the proposed work is in the combination of band selection (i.e., linear prediction (LP)-based method), spatial feature extraction (i.e., morphology profiles (MP)), and spectral transformation (i.e., principal component analysis (PCA)) to build a computationally tractable system. The preliminary result with ROSIS data shows that using the selected bands and MP features extracted from principal components (PCs) can yield the highest accuracy. We believe such finding is instructive to feature extraction/selection for spectral/spatial-based hyperspectral image classification.

Estimation of net primary productivity using multi-scale remote sensing data in Xuzhou, China

An improved Carnegie Ames Stanford Approach model based on two kinds of remote sensing data, Landsat ETM+ and MODIS, and climate variables was applied to estimate the net primary productivity (NPP) of Xuzhou in the June of 2006,2008 and 2010. The NPP of the study area decreases with the spatial scale expanding; The average NPP of terrestrial vegetation in Xuzhou shows decreasing trend in recent years because of the changes in climate and environment; The whole study area was plotted out four sub-regions, which were NPP higher sub-region, NPP high sub-region, NPP low sub-region and NPP lower sub-region. The average NPP of every sub-region was decreasing and the area percentage of lower sub-region was increasing with the scale expanding, so the NPP structure is various in different spatial scales. The NPP of the different vegetation types is significantly influenced by scale effect. In particular, the NPP of urban woodland was estimated lower value because of mixed pixel, it was increasing with the scale expanding.