Deep learning methods excel in Polarimetric SAR (PolSAR) image classification. However, existing methods typically sample an image block for each pixel with a fixed-size square window, which always contains inconsistent/incomplete content with the central pixel, resulting in many misclassifications especially in boundary and heterogeneous regions. So, a size-fixed square window is not enough for representing various terrain objects. To address this issue, we develop a content-adaptive multi-region deep network to obtain contextual consistent sampling windows for diverse terrain objects. Firstly, a complex scene of PolSAR image is partitioned into homogeneous, heterogeneous and boundary regions. Then, sampling windows with adaptive direction and scale are designed for three distinct regions. Besides, windows with central and global regions are proposed to provide additional local and global information. Finally, a fusion network is designed to adaptively combine different sampling windows to enhance classification performance. Experimental results on three real data sets demonstrate that the proposed method can achieve superior performance in both edge details and heterogeneous terrain objects compared with the state-of-the-art methods.

Diffusion model has achieved excellent performance in natural image processing, which can learn the noise distribution by the degradation and restoration processes. However, the model is limited to Gaussian noises. Actually, Polarimetric Synthetic Aperture Radar(PolSAR) images have complex non-Gaussian speckle noises, for which the Gaussian diffusion model is difficult to learn their intrinsic statistical characteristics. In this paper, we propose a novel scattering mechanism inspired non-Gaussian diffusion model for PolSAR image classification. To better simulate the PolSAR speckle noise, a mixed noise distribution is defined for PolSAR covariance matrices by combining Gamma multiplicative and Gaussian additive noises. A non-Gaussian forward noising process is derived to degrade a clean PolSAR image to a noisy image by steps. Then, the U-net structure is trained to remove noises for each step, effectively extracting non-Gaussian statistical features. However, statistical features can only characterize the overall distribution of the dataset, which is insufficient to describe complicated individual objects; the original PolSAR data reflect the detailed scattering mechanism for individual pixels, which can provide complementary object information for classification. Therefore, a scattering-statistical joint learning network is further developed with a dual-branch architecture to enhance discrimination ability. In particular, a multiscale pyramid module and attention mechanism are designed to improve the ability of feature learning. Experimental results on five real PolSAR datasets demonstrate that the proposed method effectively captures edge details and preserves homogeneous regions for terrain classification, especially in heterogeneous regions.

Deep learning models have demonstrated excellent performance for polarimetric SAR image classification. However, existing approaches generally use a fixed square window to sample image blocks as the network input, which may not effectively extract various terrain objects. To alleviate this issue, we proposed an adaptive region sampling network to learn different terrain types by introducing a novel sampling scheme with varying direction and scale. Initially, a complex PolSAR image is segmented into homogeneous, heterogeneous and boundary regions. Subsequently, small-scale and large-scale sampling windows are designed for homogeneous and heterogeneous regions, to capture local and global features for two types of regions respectively. Additionally, an adaptive directional sampling window is designed for boundary regions to ensure context consistency in the image block and prevent edge confusion. Experiments conducted on real PolSAR data sets demonstrate that our method achieves superior classification results, providing both regional consistency and boundary preservation.

Deep learning networks can automatically acquire high-level semantic features for polarimetric SAR image classification, while it involves a blind learning procedure without explicit guidance. In contrast, sparse representation methods represent effective non-deep models with a robust mathematical mechanism serving as guidance. However, they can’t capture complex image features and semantic information. To address these issues, we propose a novel approach known as the CNN-enhanced Deep Sparse Representation Network (CE-DSRNet) for PolSAR image classification, which a Sparse Representation (SR) guided deep learning model. Initially, a sparse representation model is constructed for PolSAR images to capture essential features. Subsequently, to solve the sparse model, a Deep Sparse Representation Network (DSRNet) is devised by transforming the Soft Threshold Iterative (ISTA) optimization procedure into a network, enabling automatic learning of sparse coefficients as features. Finally, a CNN-enhanced DSRNet is introduced, integrating DSRNet with CNN to effectively extract deep semantic features and enhance classification accuracy. Experiments demonstrate the effectiveness of the proposed method compared to state-of-the-art approaches.

Deep learning models have been widely applied to Polarimetric Synthetic Aperture Radar (PolSAR) image classification due to their excellent performance. However, unlike natural images, PolSAR data is a 3×3 covariance matrix for each resolution unit. Existing deep learning methods generally convert the covariance matrix into a vector as the input of neural networks, which destroys the correlation between channels and distorts the matrix structure. To alleviate this issue, we explore a Symmetric Positive Definite (SPD) convolution network for PolSAR images, which directly inputs the PolSAR complex matrix into the network to learn the geometric features in Riemannian space. Furthermore, a CNN-enhanced SPDnet is designed to further learn the contextual high-level features, which can convert Riemannian matrix features into Euclidean space and apply them for classification. Experimental results on real PolSAR data sets demonstrate the proposed method can achieve better performance than the state-of-the-art methods.

The Convolutional Neural Network (CNN) model excels at learning local features, but struggles with capturing global large-scale features, particularly in extremely heterogeneous areas. In contrast, the Graph Convolution Network (GCN) is an effective tool for PolSAR image classification, demonstrating a capability to learn large-scale global features proficiently.To learn effective features for both heterogenous terrain objects and edge details well, a novel region partition based hybrid deep network is proposed for adaptive learning features for boundary and non-boundary regions, which can learn both large-scale global features for extremely heterogeneous terrain objects and pixel-wise features for edge details. The proposed method can effectively partition a PolSAR image into boundary and non-boundary regions, and design a CNN and GCN subnetworks for them respectively. Subsequently, a unified network is designed to effectively fuse both the advantages of GCN and CNN to enhance classification performance. The experiments verify the proposed algorithm can achieve better performance than compared methods in both region homogeneity and boundary preservation.

Recently, deep learning methods have achieved superior performance for polarimetric synthetic aperture radar (PolSAR) image classification. Existing deep learning methods learn PolSAR data by converting the covariance matrix into a feature vector or complex-valued vector as the input, learning features in Euclidean space. However, it is well-known that covariance matrices are manifold data endowing in Riemannian space instead of Euclidean space. Existing methods cannot learn the geometric characteristics of covariance matrices directly and destroy the channel correlation. To learn features from covariance matrices directly, we propose a lightweight Riemannian covariance matrix convolutional network (LRCM_CNN) for PolSAR classification for the first time, which directly utilizes the covariance matrix as the network input and defines the Riemannian operations to learn complex matrix’s features in Riemannian space. The proposed LRCM_CNN network initially designs a lightweight Riemannian covariance matrix network (LRCMnet) to learn covariance matrix features by exploiting a series of Riemannian convolution, rectified linear unit (ReLu), and LogEig operations in Riemannian space, which breaks through the Euclidean constraint of conventional networks. Then, features learned from covariance matrices are converted from Riemannian to Euclidean space, and a CNN module is appended to enhance contextual covariance matrix features. Besides, a fast kernel learning method is developed for the proposed method to learn class-specific features and reduce the computation time effectively, which implements the lightweight RCMnet. Experiments are conducted on four sets of real PolSAR data with different bands and sensors. Experiments results demonstrate the proposed method can obtain superior performance than the state-of-the-art methods.

Superpixel-based graph convolution network (SGCN) can extract global features well and reduce computing time greatly, which has been widely used in image classification. However, SGCN ignores individual feature for each pixel within a superpixel. Pixel-wise convolutional neural network (CNN) can learn local features with fixed-square convolution kernel. Combine with both the advantages of SGCN and CNN, we proposed a novel SGCN-CNN method, which can combine the global and local features together. Superpixel-wise SGC-N and pixel-wise CNN cannot be combined into a network directly since they are with different scales. To alleviate this issue, encoder and decoder are designed by defining an association matrix, which can covert features between superpixel and pixel. In addition, complex matrix-based Wishart metric is used to construct the edge weights for SGCN. The proposed method can obtain both global and local features to improve classification performance. Experimental results demonstrate the effectiveness of the proposed method.

Superpixel-based graph convolutional network (SGCN) has shown the advantages of less computational time and global modeling ability for polarimetric synthetic aperture radar (PolSAR) image classification. However, the effectiveness is heavily dependent on the superpixel segmentation result. Existing superpixel segmentation methods usually produce edge errors due to speckle and scattering confusion, which directly results in the mistakes of the final classification. To address this issue, a novel hybrid weighted fuzzy SGCN method(HF-SGCN) is proposed to correct the edge pixels by defining a fuzzy projection matrix (FPM). The FPM can transform features from superpixel to pixel, by which features of edge pixels can be calculated from all the neighboring superpixels with a certain probability, so as to correct edges to the most similar region. In addition, a hybrid weighted adjacent matrix is formulated by incorporating both the revised Wishart and multi-feature distances, which can enhance the discriminating features effectively. The proposed HF-SGCN method is capable of capturing the global contextual information and rectifying edges, while disregarding the local individual features for each pixel. To combine global and local features, we further propose the HF-SGCN-CNN method, which integrates the superpixel-wise HF-SGCN network and the pixel-wise 3D-CNN network into a unified framework. Thus, we can fuse the features extracted from two subnetworks, producing complementary global and local features that significantly improve classification accuracy. Experiments are conducted on four publicly real PolSAR datasets with different sensors and bands. Experimental results demonstrate the effectiveness of the proposed methods.

Nearest-regularized subspace (NRS) algorithm is an effective tool to obtain both accuracy and speed for PolSAR image classification. However, existing NRS-based methods only use the polarimetric feature vector as the input, which cannot learn the complex matrix structure and channel information. To learn the complex matrix and scattering features collaboratively, a novel complex matrix and polarimetric feature joint learning method is proposed for PolSAR image classification. Specifically, firstly, a Riemannian NRS model is utilized to learn matrix structure by constructing complex matrix dictionary and Riemannian distance metric. Then, the complex matrix and extracted scattering features are joint learned by the proposed model by constructing coupled dictionaries and different distance metrics. Besides, superpixels are utilized to suppress the speckle noises and reduce the computing time greatly. Experiments are conducted on the real PolSAR data and the results demonstrate the effectiveness of the proposed method.

Nearest-regularized subspace (NRS) algorithm is a kind of effective representation learning method, which can obtain both accuracy and speed for PolSAR image classification. However, existing NRS methods use the polarimetric feature vector instead of the PolSAR original coherency matrix (known as Hermitian positive definite (HPD) matrix) as the input. This will destroy the matrix structure, and miss the instinct correlation among channels. How to utilize the original coherency matrix to NRS method is a key problem. To address this limitation, a Riemannian NRS method is proposed, which considers the HPD matrices endowed in a Riemannian space. First, to utilize the PolSAR original data, a Riemannian NRS method (RNRS) is proposed by constructing HPD dictionary and HPD distance metric. Then, a new Tikhonov regularization term is designed to reduce the differences within the same class. Finally, the optimization method is developed to resolve the proposed model, and the first-order derivative is inferred. Besides, only coherency matrix T is used as the input in the proposed method, while multiple features are utilized for compared methods in the experiments. Experimental results demonstrate the proposed method can outperform the state-of-the-art algorithms even with fewer features.

In conventional terrain classification for the polarimetric SAR (POLSAR) images, color features are rarely involved unless in one recent supervised work. Unlike that work, the color features are exploited for the unsupervised classification in this paper. Firstly, based on the polarimetric decomposition of the POLSAR data, the common color spaces, such as RGB, HSI, and CIELab are calculated. The color feature is quantitatively selected from these color spaces by introducing the color entropy. Then together with the spatial information, extended scattering power entropy and the copolarized ratio, the adaptive Mean-shift algorithm is used to segment the POLSAR image. Finally, the segments are merged according to the Wishart distance measurement. The experiments using AIRSAR L-band POLSAR data indicate that the proposed method has better discriminative ability for urban areas and for boundary preservation compared with existing works.