Polarimetric Synthetic Aperture Radar Image Classification Research Articles

Adaptive Fuzzy Learning Superpixel Representation for PolSAR Image Classification

The increasing applications of polarimetric synthetic aperture radar (PolSAR) image classification demand for effective superpixels’ algorithms. Fuzzy superpixels’ algorithms reduce the misclassification rate by dividing pixels into superpixels, which are groups of pixels of homogenous appearance and undetermined pixels. However, two key issues remain to be addressed in designing a fuzzy superpixel algorithm for PolSAR image classification. First, the polarimetric scattering information, which is unique in PolSAR images, is not effectively used. Such information can be utilized to generate superpixels more suitable for PolSAR images. Second, the ratio of undetermined pixels is fixed for each image in the existing techniques, ignoring the fact that the difficulty of classifying different objects varies in an image. To address these two issues, we propose a polarimetric scattering information-based adaptive fuzzy superpixel (AFS) algorithm for PolSAR images classification. In AFS, the correlation between pixels’ polarimetric scattering information, for the first time, is considered through fuzzy rough set theory to generate superpixels. This correlation is further used to dynamically and adaptively update the ratio of undetermined pixels. AFS is evaluated extensively against different evaluation metrics and compared with the state-of-the-art superpixels’ algorithms on three PolSAR images. The experimental results demonstrate the superiority of AFS on PolSAR image classification problems.

Polarimetric SAR Image Classification Based on Feature Enhanced Superpixel Hypergraph Neural Network

Synthetic aperture radar (SAR) images can capture abundant spatial and polarimetric information of land cover objects, and thus polarimetric SAR (PolSAR) image classification has been developed for various applications. Combining the advantages of spatial and polarimetric information simultaneously is of great importance for PolSAR image classification. In this paper, a feature enhanced superpixel hypergraph neural network (FESHNN) is proposed for PolSAR image classification, which aims to take full advantage of spatial features and polarimetric features from PolSAR images. In the proposed model, superpixel hypergraph neural network is constructed for feature representation of superpixels, which aims to obtain spatial correlation and polarimetric correlation in a hypergraph. Then, a feature enhancement module is employed to refine the local features of pixels and the spatial features of superpixels, which aims to enhance the discrimination of feature representation. Experimental results on three PolSAR datasets demonstrate that the proposed method yields superior classification performance compared with other related approaches.

Winner Takes All: A Superpixel Aided Voting Algorithm for Training Unsupervised PolSAR CNN Classifiers

Unsupervised methods play an essential role in Polarimetric SAR (PolSAR) image classification, where labeled data are difficult to obtain. However, there is still a large gap between existing unsupervised learning methods and supervised learning methods. Without the semantic constraints of labeled data, pixels within the same category are often misclassified into different categories, leaving the output to be messy. To address the previous issue, this paper proposes a fully unsupervised pipeline for training CNNs. The pipeline combines low-level superpixels and high-level CNN semantic features for high-quality pseudo label generation. It effectively eliminates the misclassified pixels by voting within the superpixel blob, while preserving the sharpness of edges. With the training process of the model, the quality of the generated labels is getting improved. Experiments on airborne (ESAR/AIRSAR) and spaceborne (RadarSat2) PolSAR images prove the effectiveness of the proposed method (measured with OA, AA, and Kappa metrics). Our method outperforms the previous unsupervised methods (H/alpha-Wishart,SM-Wishart, FDD-H, DEC, and VQC-CAE) with a large margin, and even has comparable performance to the supervised CNN model (FCN).

Unsupervised Complex-Valued Sparse Feature Learning for PolSAR Image Classification

Deep learning has powerful feature extraction abilities and has achieved promising results in polarimetric synthetic aperture radar (PolSAR) image classification. However, the labeled samples of PolSAR images are generally limited, which could lead to the overfitting of deep networks and the inefficiency of deep features. To overcome this problem, in this article, we propose a complex-valued enforcing population and lifetime sparsity (CV-EPLS) model to extract nonredundant sparse features from PolSAR images. CV-EPLS achieves unsupervised learning of sparse polarimetric features with limited and unlabeled samples, including amplitude and phase information in multiple polarimetric channels. Concretely, CV-EPLS defines an activation metric function to achieve strong population sparsity. Additionally, a grid search strategy is designed to ensure that activation items are evenly distributed among the sparse targets, thus forming strong lifetime sparsity. In this way, CV-EPLS constructs the complex sparse matrices and extracts discriminative sparse features in an unsupervised way, with the dependence of features being effectively reduced. Experimental results on PolSAR images demonstrate the effectiveness of CV-EPLS in the extraction of features and its application to image classification.

Adaptive Graph Convolutional Network for PolSAR Image Classification

Polarimetric synthetic aperture radar (PolSAR) image classification is one of the hottest issues in remote sensing, where studies on pixel-level information and relationship are of great significance. In this article, graph convolutional network (GCN) is employed to accomplish this pixel-level task benefiting from its excellent capability in structure exploration and information propagation between different pixels. To reduce the communication burden between various PolSAR pixels and high computational cost for the whole PolSAR image, an adaptive GCN (AdapGCN) consisting of pixel-centered subgraphs is proposed in this article. In the AdapGCN, a data-adaptive kernel and a spatial-adaptive kernel are introduced to, respectively, model data structure and spatial structure for PolSAR image. Moreover, a multiscale learning structure is integrated to further explore complicated relations between pixels. Extensive comparative evaluations validate the superiority of our new AdapGCN model for PolSAR image classification over a wide range of state-of-the-art methods on three challenging benchmarks.

Complex Matrix and Multi-Feature Collaborative Learning for Polarimetric Sar Image Classification

Complex Matrix and Multi-Feature Collaborative Learning for Polarimetric Sar Image Classification

Texture descriptors for robust SAR image segmentation

SAR (synthetic aperture radar) and PolSAR (polarimetric synthetic aperture radar) images fulfill a fundamental role in Earth observation, due to their advantages over optical images. However, the presence of speckle noise hinders their automatic interpretation and unsupervised use, rendering traditional segmentation tools ineffective. For this reason, advanced image segmentation models are sought to overcome the limitations that make an adequate treatment of speckled images difficult. We propose a procedure for SAR and PolSAR image classification, based on texture descriptors, that combines fractal dimension, a specific probability distribution function, Tsallis entropy, and the entropic index. A vector of local texture features is built using a set of reference regions, then a support vector machine classifier is applied. The proposed algorithm is tested with synthetic and actual monopolarimetric and polarimetric SAR imagery, exhibiting visually remarkable and robust results in coincidence with quantitative quality metrics as accuracy and F1-score.

Deep Triplet Complex-Valued Network for PolSAR Image Classification

Recently, convolutional neural network (CNN) has proved itself as a successful deep model and has been successfully utilized in polarimetric synthetic aperture radar (PolSAR) image classification. Most CNN-based models, however, concentrate on the correlation between the pixels and the labels in images and have fewer constraints on interclass or intraclass features. For fully utilizing the polarimetric data, we utilized complex-valued (CV) distance to learn the PolSAR features and proposed a PolSAR classification method by CV distance comparisons. First, we proposed a triplet CV network (TCVN) to learn the CV representations from PolSAR data by maximizing the interclass distance and minimizing intraclass distance. It uses the CV convolution and the CV Euclidean to maintain the phase components and applies the CV-dropout and CV <inline-formula> <tex-math notation="LaTeX">$L_{2}$ </tex-math></inline-formula> parameter regularization to reduce the overfitting and further improve the network performance. Subsequently, CV K nearest neighbor (CV-KNN) computes the distance of the CV representations and groups similar pixels. CV-KNN is well coupled with the TCVN because both of them are based on the Euclidean distance in the complex domain. Compared with the CNN-based methods, the proposed deep metric learning model can simultaneously extract the hierarchical features by comparing the polarimetric resolution cells in the complex domain and maintain the phase component by performing CV convolutions. The effectiveness and the superiorities of CV Euclidean distance in TCVN are demonstrated. Experiments on real PolSAR images illustrate that TCVN can deal with PolSAR data more effectively and achieve comparable performance in the PolSAR image classification even with a smaller data set.

Convolutional Kernel‐based covariance descriptor for classification of polarimetric synthetic aperture radar images

There are two types of important information in a polarimetric synthetic aperture radar (PolSAR) image: spatial features in two dimensions and polarimetric characteristics in the scattering dimension. Considering both polarimetric and spatial information is important for PolSAR image classification. Convolutional kernels show superior performance for extraction of spatial information from two dimensions of an image in convolutional neural networks (CNNs). But learning CNNs needs large enough training sets to achieve the optimum weights of kernels while there are not usually sufficient training samples for PolSAR images. To deal with this difficulty, a convolutional kernel-based covariance descriptor (CKCD) is introduced for PolSAR image classification in this study. To extract contextual characteristics, compatible with the original image, the fixed-valued convolutional kernels randomly selected from the image are used, which do not require any learning, and so do not need any training samples. To include more local spatial information and find the relation among the polarimetric features, the covariance descriptor is constructed on the extracted feature maps. Then, the polarimetric-contextual features are given to a support vector machine with a matrix logarithm-based kernel. Finally, the guided filter is applied to the initial classification map to result a smoothed classification map with preserved edges. The experiments on three real PolSAR images show superiority of the proposed CKCD method compared to several PolSAR classification methods such as 2DCNN and 3DCNN in small sample size situations.

Terrain Classification of Polarimetric SAR Images Using Semi-supervised Spatial-channel Selective Kernel Network

In this paper, a Spatial-Channel Selective Kernel Fully Convolutional Network (SCSKFCN) and a Semi-supervised Preselection-United Optimization (SPUO) method are proposed for polarimetric Synthetic Aperture Radar (SAR) image classification. Integrated with spatial-channel attention mechanism, SCSKFCN adaptively fuses features that have different sizes of reception field, and achieves promising classification performance. SPUO can efficiently extract information contained in unlabeled samples according to annotated samples. It utilizes K-Wishart distance to preselect unlabeled samples for pseudo label generation, and then optimizes SCSKFCN with both labeled and pseudo labeled samples. During the training process of SCSKFCN, a two-step verification mechanism is applied on pseudo labeled samples to reserve reliable samples for united optimization. The experimental results show that the proposed SCSKFCN-SPUO can achieve promising performance and efficiency using limited number of annotated pixels.

Integration of the k-nearest neighbours and patch-based features for PolSAR image classification by using a two-branch residual network

ABSTRACT A two-branch residual deep learning network is proposed for polarimetric synthetic aperture radar (PolSAR) image classification in this work. The proposed method integrates local information of pixels from a rectangular region (patch-based features) with global information of the k-nearest neighbours (k-NN) constituting a deformable shape. Both contextual and polarimetric information are utilized to find the nearest neighbours. The experiments on two real Airborne Synthetic Aperture Radar (AIRSAR) datasets show superior performance of the proposed network with a relatively small training set.

Deep support vector machine for PolSAR image classification

ABSTRACT The main problem posed by Polarimetric Synthetic Aperture Radar (PolSAR) image classification in remote sensing is the ability to develop classifiers that can substantially discern the different classes inherent in natural and man-made targets. Emphasis has shifted from the use of conventional classifiers to modern non-parametric classifiers such as the Artificial Neural Network (ANN) and Support Vector Machine (SVM), and most recently the hybrid Deep Neural Network (DNN) which is a fusion of Deep Learning (DL) and ANN. This research therefore presents the novel application of Deep Support Vector Machine (DSVM), which is a fusion of DL and SVM to PolSAR image classification. Two PolSAR images of Flevoland region in the Netherlands and Winnipeg in Canada are used as test beds for the experiment. The Lee filter is used to filter the images to suppress the speckle noise in the images. The Pauli decomposition is applied to decompose the images into , , polarimetric channels. Then, the Gray Level Co-occurrence Matrix (GLCM) texture feature for , , are extracted based on correlation, contrast, energy, and homogeneity statistics, using GLCM directions 0°, 45°, 90°, and 135° with an offset distance of 60. To enhance the efficiency of the model 8, 16, 32, 64, 128, and 256 quantization levels are explored. The DSVM classifier is implemented with four kernel functions: Exponential Radial Basis Function (ERBF), Gaussian Radial Basis Function (GRBF), neural, and polynomial. The first set of results is a comparison of the DSVM and SVM. The result of Flevoland image for ERBF, GRBF, neural, and polynomial kernels are 99.17 (73.39), 99.32 (74.62), 98.64 (71.28), and 99.34 (77.21), respectively; for the Winnipeg image for ERBF, GRBF, neural, and polynomial kernels are 98.65 (72.68), 98.67 (73.54), 98.27 (70.15), and 99.46 (75.03), respectively. The second set of results is a comparison of DSVM, SVM, DNN, Gaussian Mixture Model (GMM), K Nearest Neighbour (KNN), and K Means (KM) classifiers; the results for Flevoland image for DSVM, SVM, DNN, GMM, KNN, and KM are 99.12, 74.13, 96.29, 75.06, 75.85, and 21.43, respectively, while the results for Winnipeg image for DSVM, SVM, DNN, GMM, KNN, and KM are 98.76, 72.85, 95.64, 73.20, 73.91, and 25.60, respectively. Since the Kappa coefficient is presumed to be a more accurate measure for accuracy estimation, it is used to evaluate the performances of all the models. The computed Kappa coefficients for of DSVM, SVM, DNN, GMM, KNN, and KM for Flevoland are 92.45, 70.71, 88.76, 68.59, 68.62, and 18.89, respectively; while the computed Kappa coefficients for DSVM, SVM, DNN, GMM, KNN, and KM for Winnepeg are 92.45, 70.71, 88.76, 68.59, 68.62, and 18.89, respectively. Based on the metrics used to evaluate the performances of the experiments; the results show that the DSVM outperformed the other classifiers. The high accuracy obtained with the DSVM shows it is a state-of-the-art algorithm for PolSAR image classification and a significant progress in the latest of DL applications.

PolSAR Image Land Cover Classification Based on Hierarchical Capsule Network

Polarimetric synthetic aperture radar (PolSAR) image classification is one of the basic methods of PolSAR image interpretation. Deep learning algorithms, especially convolutional neural networks (CNNs), have been widely used in PolSAR image classification due to their powerful feature learning capabilities. However, a single neuron in the CNN cannot represent multiple polarimetric attributes of the land cover. The capsule network (CapsNet) uses vectors instead of the single neuron to characterize the polarimetric attributes, which improves the classification performance compared with traditional CNNs. In this paper, a hierarchical capsule network (HCapsNet) is proposed for the land cover classification of PolSAR images, which can consider the deep features obtained at different network levels in the classification. Moreover, we adopt three attributes to uniformly describe the scattering mechanisms of different land covers: phase, amplitude, and polarimetric decomposition parameters, which improves the generalization performance of HCapsNet. Furthermore, conditional random field (CRF) is added to the classification framework to eliminate small isolated regions of the intra-class. Comprehensive evaluations are performed on three PolSAR datasets acquired by different sensors, which demonstrate that our proposed method outperforms other state-of-the-art methods.

Nonstationary PolSAR Image Classification by Deep-Features-Based High-Order Triple Discriminative Random Field

Aiming at exploiting the discriminative deep features and encoding the high-level structures, this letter presents a deep-features-based high-order triple discriminative random field model, abbreviated as DF-HoTDF, for nonstationary polarimetric synthetic aperture radar (PolSAR) image classification. First, the DF-HoTDF model extracts the discriminative deep features by a graph-based complex-valued 3-D convolutional neural network (CV-3-D-CNN) and then constructs the unary potential by a negative log function. Second, it introduces an auxiliary field u to explicitly regulate the nonstationary label patterns of the PolSAR image and then constructs a pairwise potential guided by u to capture greater pairwise label interactions. Third, it defines a high-order potential on high-order cliques to encode high-level structures. Finally, under the discriminative model framework, the DF-HoTDF model has a weighted fusion of the unary potential, the pairwise potential, and the high-order potential. Then, with the DF-HoTDF model, we iteratively optimize the class label and the stationary maps until they converge. The experimental results demonstrate that the proposed DF-HoTDF model is of superior performances in nonstationary PolSAR image classification and that it can provide better label consistency in homogeneous region and better target structures and edge locations in heterogeneous region.

Statistical Scattering Component-Based Subspace Alignment for Unsupervised Cross-Domain PolSAR Image Classification

Increasing amounts of polarimetric synthetic aperture radar (PolSAR) images from different sensors covering different scenes are available, but limited labeled samples and trained models can hardly work well in these cross-domain data interpretations. Fortunately, domain adaptation (DA) can transfer knowledge in existing images to new yet related images. DA shows attractive potential for PolSAR classification, and it is still challenging due to more complex domain shifts caused by different sensors, imaging conditions, and distributions. Inspired by the widely applicable polarimetric scattering mechanisms and DA ability of subspace alignment (SA), this article is devoted to constructing a robust unsupervised cross-domain PolSAR classification framework, by exploring scattering and statistical characteristics mapping between the source and target domains. First, classical scattering components of both source and target data were extracted, and Wishart clustering was adopted to derive the statistical information of scattering components at patch level. Second, the intrinsic polarimetric scattering components were estimated and extracted, which were called statistical scattering components (SSCs). Third, by applying SA, the source SSC was aligned with target SSC, and domain shift was further reduced. Finally, the target PolSAR image was classified based on labeled samples from source domain, and unsupervised cross-domain classification was achieved by SSC-based SA (SSC-SA). The unsupervised cross-domain experiments are conducted on 49 units among 11 data sets, including Radarsat-2, Gaofen-3, AIRSAR, and Pi-SAR images. With randomly selected labeled samples (about 2&#x0025;&#x2013;10&#x0025;) from source domain, the accuracies of the proposed cross-domain classifications range between 80.20&#x0025; and 95.64&#x0025;. Also, the proposed SSC feature pattern is proved extensible for other polarimetric basis and decompositions.

Polarimetric Multipath Convolutional Neural Network for PolSAR Image Classification

Scatter targets of complex land covers in polarimetric synthetic aperture radar (PolSAR) images are often randomly oriented and cause randomly fluctuating echoes, which brings a challenge to PolSAR image classification. Therefore, many existing methods have alleviated this problem through orientation compensation. However, there are still two obstacles that limit the improvement of classification accuracy. On the one hand, generally, these methods process PolSAR images with fixed polarization rotation angles, which is experience-dependent and inflexible. On the other hand, for the different land covers of a PolSAR image, the existing methods do not consider these rotation angles separately. For the first obstacle, we design a group of convolution kernels called polarization rotation kernels (PRKs) and utilize them to build the polarimetric convolutional neural network (CNN) (PolCNN). The PolCNN is the base network of our final model, and it can learn polarization rotation angles adaptively. For the second obstacle, we extend the PolCNN into a multipath structure, the final model polarimetric multipath CNN (PolMPCNN). The polarization rotation angles of different land covers are directly related to the networks of different paths within the PolMPCNN. Furthermore, we also put forward the two-scale sampling and the stagewise training algorithm in order that our PolMPCNN can fit different scales of PolSAR targets and pays more attention to difficult training samples. Experiments on real PolSAR images show that the proposed model achieves the best classification results with an extremely low sampling rate of 0.1%.

A General Feature Paradigm for Unsupervised Cross-Domain PolSAR Image Classification

Limited labels and increasing multisource data promote domain adaptation (DA) problem as a challenging study for polarimetric synthetic aperture radar (PolSAR) interpretation. Existing DAs for optical images cannot generalize over PolSAR imagery due to its special side-imaging characteristics and complex distribution shifts. In this letter, a general feature paradigm (GFP) is proposed for unsupervised cross-domain PolSAR image classification. The GFP is based on a key observation that interclass aggregation is optimized after four-step feature transformations. This key observation leads to GFP that not only reduces the domain shifts but also compatible with typical DA methods. The GFPs are conducted on both source and target domain by unsupervised manner, including polarimetric basis extraction, the Wishart clustering, histogram statistics, and dimensionality reduction. After these transformations, the unlabeled target PolSAR image can be classified based on obtained GFP, DA, and limited labeled samples only from the source domain. Extensive unsupervised cross-domain experiments on 27 scenarios verified that GFP leads to at most 93.76% accuracy for full- and dual-polarized synthetic aperture radar (SAR) images’ classification. Moreover, the GFP shed light on extensive cross-domain PolSAR applications about built-up areas, vegetation, and bare land analysis.

Multiple Classifiers Based Semi-Supervised Polarimetric SAR Image Classification Method.

Polarimetric synthetic aperture radar (PolSAR) image classification has played an important role in PolSAR data application. Deep learning has achieved great success in PolSAR image classification over the past years. However, when the labeled training dataset is insufficient, the classification results are usually unsatisfactory. Furthermore, the deep learning approach is based on hierarchical features, which is an approach that cannot take full advantage of the scattering characteristics in PolSAR data. Hence, it is worthwhile to make full use of scattering characteristics to obtain a high classification accuracy based on limited labeled samples. In this paper, we propose a novel semi-supervised classification method for PolSAR images, which combines the deep learning technique with the traditional scattering trait-based classifiers. Firstly, based on only a small number of training samples, the classification results of the Wishart classifier, support vector machine (SVM) classifier, and a complex-valued convolutional neural network (CV-CNN) are used to conduct majority voting, thus generating a strong dataset and a weak dataset. The strong training set are then used as pseudo-labels to reclassify the weak dataset by CV-CNN. The final classification results are obtained by combining the strong training set and the reclassification results. Experiments on two real PolSAR images on agricultural and forest areas indicate that, in most cases, significant improvements can be achieved with the proposed method, compared to the base classifiers, and the improvement is approximately 3–5%. When the number of labeled samples was small, the superiority of the proposed method is even more apparent. The improvement for built-up areas or infrastructure objects is not as significant as forests.

Representative Learning via Span-Based Mutual Information for PolSAR Image Classification

The optimal parameters of polarimetric scattering decomposition are critical to classify the pixels in polarimetric synthetic aperture radar (PolSAR) images by utilizing the method of machine learning. Therefore, span-based mutual information (Sp-MI) is proposed to lighten the dependence on labeling information, and then a heuristic representative learning scheme is also given by artificial neural network (ANN) to classify parameters separately with the increasing sequence according to the values of Sp-MI. Furthermore, an innovative method of using the sine function is presented to map the parameters of angular, and a min-max scaling method is applied to complete the procedure of normalization. Except for the support vector machine, three ANN-based classifiers are implemented to verify the rationality and effectiveness of the proposed representative learning and normalization scheme. Meanwhile, the classification method is compared with four similar comparison methods on three real PolSAR images. Finally, the classification results show the effectiveness of the proposed Sp-MI and the validation of the representative learning scheme in the aspect of classification overall accuracy and visual effect.

Graph Convolutional Networks by Architecture Search for PolSAR Image Classification

Classification of polarimetric synthetic aperture radar (PolSAR) images has achieved good results due to the excellent fitting ability of neural networks with a large number of training samples. However, the performance of most convolutional neural networks (CNNs) degrades dramatically when only a few labeled training samples are available. As one well-known class of semi-supervised learning methods, graph convolutional networks (GCNs) have gained much attention recently to address the classification problem with only a few labeled samples. As the number of layers grows in the network, the parameters dramatically increase. It is challenging to determine an optimal architecture manually. In this paper, we propose a neural architecture search method based GCN (ASGCN) for the classification of PolSAR images. We construct a novel graph whose nodes combines both the physical features and spatial relations between pixels or samples to represent the image. Then we build a new searching space whose components are empirically selected from some graph neural networks for architecture search and develop the differentiable architecture search method to construction our ASGCN. Moreover, to address the training of large-scale images, we present a new weighted mini-batch algorithm to reduce the computing memory consumption and ensure the balance of sample distribution, and also analyze and compare with other similar training strategies. Experiments on several real-world PolSAR datasets show that our method has improved the overall accuracy as much as 3.76% than state-of-the-art methods.