In existing superpixel-wise segmentation algorithms, superpixel generation most often is an isolated preprocessing step. The segmentation performance is determined to a certain extent by the accuracy of superpixels. However, it is still a challenge to develop a stable superpixel generation method. In this article, we attempt to incorporate the superpixel generation and merging steps into an end-to-end trainable deep network. First, we employ a recently proposed differentiable superpixel generation method to over-segment the single-polarization synthetic aperture radar (SAR) image. It outputs the statistical likelihood that each pixel belongs to different superpixels. In superpixel merging part, as one of our main contributions, we propose a superpixel-wise statistical dissimilarity measure method for converting the soft superpixels set into a self-connected weighted graph. More importantly, inspired by the concept of the number of walks in graph theory, we define the $k$ -order connectivity of each vertex. This definition can intelligently indicate the potential soft cluster centers and class assignments in graph. This merging method is differentiable, computationally simple, and free of empirical parameters. The superpixel generation and merging phases can be implemented under a unified deep network. The benefit is that our method can iteratively adjust the shapes of the superpixels according to the boundaries and segmentation results during training, until the satisfactory segmentation results are captured. Experimental results on real SAR images demonstrate that the segmentation precision of our proposed method is superior to other state-of-the-art methods in terms of precision and computational efficiency.

Gaofen-3 is China’s first meter-level multipolarization synthetic aperture radar (SAR) satellite with 12 imaging modes for the scientific and commercial applications. In order to evaluate the imaging performance of these modes, the multiple mode SAR raw data simulation is highly demanded. In the paper, the multiple mode SAR simulation framework will be briefly introduced to expose how the raw data simulation guarantees the development of Gaofen-3 and its ground processing system. As an engineering simulation, the complex working modes and practical evaluation requirements of Gaofen-3 mission put forward to the higher demand for simulation simplification and data input/output (I/O) efficiency. To meet the requirements, two improvements have been proposed. First, the stripmap mode based multiple mode decomposition method is introduced to make a solid and simplified system simulation structure. Second, the cloud computing and graphics processing unit (GPU) are integrated to simulate the practical huge volume raw data, resulting in improved calculation and data I/O efficiency. The experimental results of sliding spotlight imaging prove the effectiveness of the Gaofen-3 mission simulation framework and the decomposition idea. The results for efficiency assessment show that the GPU cloud method greatly improves the computing power of a 16-core CPU parallel method about $40\times$ speedup and the data throughput with the Hadoop distributed file system. These results prove that the simulation system has the merits of coping with multiple modes and huge volume raw data simulation and can be extended to the future space-borne SAR simulation.

Since the number of superpixels is lower than that of pixels, superpixels can substantially speed up subsequent processing steps and have been widely used in synthetic aperture radar (SAR) image segmentation. However, in most of the existing superpixel-wise segmentation algorithms, superpixel prediction is an isolated preprocessing step and is independent of the segmentation task. The performance of the segmentation results is determined by the accuracy of superpixels. Once superpixels are generated, their shape cannot be changed in the following segmentation stage, even if the same superpixels contain pixels of different landcovers. To address this, we propose an end-to-end trainable superpixel-wise segmentation method for single-polarization SAR images. First, we design a differentiable boundary-ware clustering method for estimating task-specific superpixels. Instead of the hard association between pixels and superpixels in the existing superpixel algorithms, this method introduces the soft association map to make the clustering differentiable. Hence, it can be implemented using a simple deep fully convolutional network. In the segmentation part, we propose a novel soft graph convolution network (Soft-GCN), which takes the association map as input and performs superpixel-wise segmentation. The advantage of our method is that superpixel generation and graph convolution parts can be trained under a unified framework, until two parts obtain the optimum parameters. In the training process, it can adaptively adjust the shape of the superpixels according to the segmentation results, ensuring the superpixels correctly adhere the boundaries. Experimental results with simulated and real SAR images demonstrate that our method outperforms other state-of-the-art segmentation algorithms, while also being faster.

Convolutional neural networks (CNNs) have demonstrated impressive ability to achieve promising results in PolSAR image classification. However, the traditional CNN performs convolution on local square regions with fixed sizes. The selection of these local square regions (patches) cannot fully take advantage of the boundary information of land covers and cannot search optimal neighborhoods in the whole image. To overcome these shortcomings, we propose a superpixel-based graph convolutional network (SP-GCN) for PolSAR image classification. SP-GCN utilizes superpixels as graph nodes, which makes full use of boundary information of superpixels and significantly reduces the computational cost of GCN, making it possible to apply GCN to large-scale PolSAR image classification. To reduce the impact of superpixel scale on classification results, we further propose a multiscale superpixel-based graph convolutional network (MSSP-GCN) based on the SP-GCN. Experimental results on three PolSAR datasets firmly demonstrate the superiority of the proposed SP-GCN and MSSP-GCN to other state-of-the-art methods.

Deep convolutional neural networks (CNN) have been recently applied to synthetic aperture radar (SAR) for automatic target recognition (ATR) and have achieved state-of-the-art results with significantly improved recognition performance. However, the training period of deep CNN is long, and the size of the network is huge, sometimes reaching hundreds of megabytes. These two factors of deep CNN hinders its practical implementation and deployment in real-time SAR platforms that are typically resource-constrained. To address this challenge, this paper presents three strategies of network compression and acceleration to decrease computing and memory resource dependencies while maintaining a competitive accuracy. First, we introduce a new weight-based network pruning and adaptive architecture squeezing method to reduce the network storage and the time of inference and training process, meanwhile maintain a balance between compression ratio and classification accuracy. Then we employ weight quantization and coding to compress the network storage space. Due to the fact that the amount of calculation is mainly reflected in the convolution layer, a fast approach for pruned convolutional layers is proposed to reduce the number of multiplication by exploiting the sparsity in the activation inputs and weights. Experimental results show that the convolutional neural networks for SAR-ATR can be compressed by $40\times$ without loss of accuracy, and the number of multiplication can be reduced by $15\times$ . Combining these strategies, we can easily load the network in resource-constrained platforms, speed up the inference process to get the results in real-time or even retrain a more suitable network with new image data in a specific situation.

Polarimetric features of PolSAR images include inherent scattering mechanisms of terrain types, which are important for classification and other Earth observation applications. By using target decomposition methods, many polarimetric scattering components can be obtained. Besides, the elements of a coherency/covariance matrix, as well as polarimetric descriptors, such as SPAN, single-bounce eigenvalue relative difference/double-bounce eigenvalue relative difference, etc., can also provide characteristic information. In fact, more and more polarimetric decomposition components and descriptors have been proposed; the computation cost increases if all of them are employed as the input of the classification process. Although all these features obtained from the coherency/covariance matrix are not independent, still, finding out which ones are significant for the classification of different terrain types will improve the understanding of scattering mechanisms. In this article, the effective polarimetric feature combination is studied based on the vegetation classification performance of support vector machine (SVM) and nearest-regularized subspace (NRS) machine learning approaches, as well as their combinations with a Markov random field (MRF). A framework on the basis of similarity and the orthogonal subspace projection (OSP) method in a hyperspectral area is used to select the polarimetric features. For the airborne PolSAR data in Flevoland, The Netherlands, 107 polarimetric features are extracted, including matrix elements, target decomposition components, and polarimetric descriptors. A subset is selected by using the proposed and OSP methods. They have a good classification accuracy evaluated by SVM+MRF and NRS+MRF classifiers. However, when the SVM and the NRS are used without combining spatial information of the MRF, the features selected by the proposed framework with correlation coefficient criteria have much better classification performance than those of OSP and principal component analysis.