Synthetic Aperture Radar Classification Research Articles

Research on Input Schemes for Polarimetric SAR Classification Using Deep Learning

This study employs the reflection symmetry decomposition (RSD) method to extract polarization scattering features from ground object images, aiming to determine the optimal data input scheme for deep learning networks in polarimetric synthetic aperture radar classification. Eight distinct polarizing feature combinations were designed, and the classification accuracy of various approaches was evaluated using the classic convolutional neural networks (CNNs) AlexNet and VGG16. The findings reveal that the commonly employed six-parameter input scheme, favored by many researchers, lacks the comprehensive utilization of polarization information and warrants attention. Intriguingly, leveraging the complete nine-parameter input scheme based on the polarization coherence matrix results in improved classification accuracy. Furthermore, the input scheme incorporating all 21 parameters from the RSD and polarization coherence matrix notably enhances overall accuracy and the Kappa coefficient compared to the other seven schemes. This comprehensive approach maximizes the utilization of polarization scattering information from ground objects, emerging as the most effective CNN input data scheme in this study. Additionally, the classification performance using the second and third component total power values (P2 and P3) from the RSD surpasses the approach utilizing surface scattering power value (PS) and secondary scattering power value (PD) from the same decomposition.

An Ensemble Approach using Self-attention based MobileNetV2 for SAR classification

The Synthetic Aperture Radar (SAR) image classification has become an essential task in numerous applications. Despite its significance, SAR image classification remains challenging due to the inherent complexity and speckle noise characteristics of SAR data. To address this, an innovative approach is proposed in this study, combining the power of ensemble learning and the efficiency of MobileNetV2 architecture to achieve robust and accurate SAR image classification. The proposed deep learning model consists of two main components: an ensemble learning approach that combines multiple binary classifiers, and the employment of MobilenetV2 architecture with self-attention mechanism as a binary classifier. The ensemble approach uses a max-value selector for predicting final class from the individual classifier output. The MobileNetV2 architecture is finetuned by freezing weights and removing the last four layers. The inclusion of the attention layer within the architecture aids in refining the learning process by focusing on crucial features while disregarding irrelevant background elements in the images.. The effectiveness of the model is evaluated using Moving and Stationary Target Acquisition and Recognition (MSTAR) benchmark dataset with accuracy of classification as performance metric.The model is trained and tested with images captured at various aspect angles. This approach enhances the robustness of the results, making them more applicable to real-world scenarios when compared to prior methods. This study also compares the performance of the ensemble method using MobileNetV2 with state-of-art individual classifiers and different CNN architectures. The proposed ensemble method OVA-MbNN achieves an impressive classification accuracy of 76.63%, surpassing the performance of state-of-the-art classifiers while maintaining robustness.

Land Cover Classification of SAR Based on 1DCNN-MRF Model Using Improved Dual-Polarization Radar Vegetation Index

Accurate land cover classification (LCC) is essential for studying global change. Synthetic aperture radar (SAR) has been used for LCC due to its advantage of weather independence. In particular, the dual-polarization (dual-pol) SAR data have a wider coverage and are easier to obtain, which provides an unprecedented opportunity for LCC. However, the dual-pol SAR data have a weak discrimination ability due to limited polarization information. Moreover, the complex imaging mechanism leads to the speckle noise of SAR images, which also decreases the accuracy of SAR LCC. To address the above issues, an improved dual-pol radar vegetation index based on multiple components (DpRVIm) and a new LCC method are proposed for dual-pol SAR data. Firstly, in the DpRVIm, the scattering information of polarization and terrain factors were considered to improve the separability of ground objects for dual-pol data. Then, the Jeffries-Matusita (J-M) distance and one-dimensional convolutional neural network (1DCNN) algorithm were used to analyze the effect of difference dual-pol radar vegetation indexes on LCC. Finally, in order to reduce the influence of the speckle noise, a two-stage LCC method, the 1DCNN-MRF, based on the 1DCNN and Markov random field (MRF) was designed considering the spatial information of ground objects. In this study, the HH-HV model data of the Gaofen-3 satellite in the Dongting Lake area were used, and the results showed that: (1) Through the combination of the backscatter coefficient and dual-pol radar vegetation indexes based on the polarization decomposition technique, the accuracy of LCC can be improved compared with the single backscatter coefficient. (2) The DpRVIm was more conducive to improving the accuracy of LCC than the classic dual-pol radar vegetation index (DpRVI) and radar vegetation index (RVI), especially for farmland and forest. (3) Compared with the classic machine learning methods K-nearest neighbor (KNN), random forest (RF), and the 1DCNN, the designed 1DCNN-MRF achieved the highest accuracy, with an overall accuracy (OA) score of 81.76% and a Kappa coefficient (Kappa) score of 0.74. This study indicated the application potential of the polarization decomposition technique and DEM in enhancing the separability of different land cover types in SAR LCC. Furthermore, it demonstrated that the combination of deep learning networks and MRF is suitable to suppress the influence of speckle noise.

Fusion and Classification of SAR and Optical Data Using Multi-Image Color Components with Differential Gradients

This paper proposes a gradient-based data fusion and classification approach for Synthetic Aperture Radar (SAR) and optical image. This method is used to intuitively reflect the boundaries and edges of land cover classes present in the dataset. For the fusion of SAR and optical images, Sentinel 1A and Sentinel 2B data covering Central State Farm in Hissar (India) was used. The major agricultural crops grown in this area include paddy, maize, cotton, and pulses during kharif (summer) and wheat, sugarcane, mustard, gram, and peas during rabi (winter) seasons. The gradient method using a Sobel operator and color components for three directions (i.e., x, y, and z) are used for image fusion. To judge the quality of fused image, several fusion metrics are calculated. After obtaining the resultant fused image, gradient based classification methods, including Stochastic Gradient Descent Classifier, Stochastic Gradient Boosting Classifier, and Extreme Gradient Boosting Classifier, are used for the final classification. The classification accuracy is represented using overall classification accuracy and kappa value. A comparison of classification results indicates a better performance by the Extreme Gradient Boosting Classifier.

Uncertainty-Incorporated Ice and Open Water Detection on Dual-Polarized SAR Sea Ice Imagery

Algorithms designed for ice–water classification of synthetic aperture radar (SAR) sea ice imagery produce only binary (ice and water) output typically using manually labeled samples for assessment. This is limiting because only a small subset of labeled samples are used, which, given the nonstationary nature of the ice and water classes, will likely not reflect the full scene. To address this, we implement a binary ice–water classification in a more informative manner considering the uncertainty associated with each pixel in the scene. To accomplish this, we have implemented a Bayesian convolutional neural network (CNN) with variational inference to produce both aleatoric (data-based) and epistemic (model-based) uncertainty. This valuable information provides feedback as to regions that have pixels more likely to be misclassified and provides improved scene interpretation. Testing was performed on a set of 21 RADARSAT-2 dual-polarization SAR scenes covering a region in the Beaufort Sea captured regularly from April to December. The model is validated by demonstrating: 1) a positive correlation between misclassification rate and model uncertainty and 2) a higher uncertainty during the melt and freeze-up transition periods, which are more challenging to classify. By incorporating the iterative region growing with semantics (IRGS) segmentation algorithm and an uncertainty value-based thresholding algorithm, the Bayesian CNN classification outputs are improved significantly via both numerical analysis and visual inspection.

Complex-Valued End-to-End Deep Network With Coherency Preservation for Complex-Valued SAR Data Reconstruction and Classification

Deep learning models have achieved remarkable success in many different fields and attracted many interests. Several researchers attempted to apply deep learning models to Synthetic Aperture Radar (SAR) data processing, but it did not have the same breakthrough as the other fields, including optical remote sensing. SAR data are in complex domain by nature and processing them with Real-Valued (RV) networks neglects the phase component which conveys important and distinctive information. A Complex-Valued (CV) end-to-end deep network is developed in this study for the reconstruction and classification of CV-SAR data. Azimuth subaperture decomposition is utilized to incorporate physics-aware attributes of the CV-SAR into the deep model. Moreover, the correlation coefficient amplitude (Coherence) of the CV-SAR images depends on the SAR system characteristics and physical properties of the target. This coherency should be considered and preserved in the processing chain of the CV-SAR data. The coherency preservation of the CV deep networks for CV-SAR images, which is mostly neglected in the literature, is evaluated in this study. Furthermore, a large-scale CV-SAR annotated dataset for the evaluation of the CV deep networks is lacking. A semantically annotated CV-SAR dataset from Sentinel-1 Single Look Complex StripMap mode data (S1SLC_CVDL dataset) is developed and introduced in this study. The experimental analysis demonstrated the better performance of the developed CV deep network for CV-SAR data classification and reconstruction in comparison to the equivalent RV model and more complicated RV architectures, as well as its coherency preservation and physics-aware capability.

Phase Bias Estimation and Imaging for High-Squint Multichannel SAR in Azimuth

High-squint multichannel synthetic aperture radar (SAR) in azimuth can overcome the constraint of minimum antenna area in the traditional single-channel classical SAR system, and thereby it would be much more flexible to achieve high-resolution and wide-swath (HRWS) imaging than broadside SAR. However, for the high-squint multichannel SAR imaging mode, the squint angle causes Doppler centroid varying with the range frequency, which can further result in much more complicated problem of the spectral folding after down-sampled with low PRF. Meanwhile, the Doppler centroid frequency is indispensable for the estimation of phase bias and unambiguous signal reconstruction among received channels. In this paper, the impact of high-squint angle on the aliased Doppler spectrum and Doppler centroid is analyzed in detail. Then, a novel approach to phase bias estimation as well as the underlying signal processing in the case of high-squint multichannel SAR imaging mode in azimuth is proposed. Simulation results validate the effectiveness of the proposed processing flowchart for high-squint multichannel SAR.

RBFA-Net: A Rotated Balanced Feature-Aligned Network for Rotated SAR Ship Detection and Classification

Ship detection with rotated bounding boxes in synthetic aperture radar (SAR) images is now a hot spot. However, there are still some obstacles, such as multi-scale ships, misalignment between rotated anchors and features, and the opposite requirements for spatial sensitivity of regression tasks and classification tasks. In order to solve these problems, we propose a rotated balanced feature-aligned network (RBFA-Net) where three targeted networks are designed. They are, respectively, a balanced attention feature pyramid network (BAFPN), an anchor-guided feature alignment network (AFAN) and a rotational detection network (RDN). BAFPN is an improved FPN, with attention module for fusing and enhancing multi-level features, by which we can decrease the negative impact of multi-scale ship feature differences. In AFAN, we adopt an alignment convolution layer to adaptively align the convolution features according to rotated anchor boxes for solving the misalignment problem. In RDN, we propose a task decoupling module (TDM) to adjust the feature maps, respectively, for solving the conflict between the regression task and classification task. In addition, we adopt a balanced L1 loss to balance the classification loss and regression loss. Based on the SAR rotation ship detection dataset, we conduct extensive ablation experiments and compare our RBFA-Net with eight other state-of-the-art rotated detection networks. The experiment results show that among the eight state-of-the-art rotated detection networks, RBFA-Net makes a 7.19% improvement with mean average precision compared to the second-best network.

The Use of C-Band and X-Band SAR with Machine Learning for Detecting Small-Scale Mining

Illicit small-scale mining occurs in many tropical regions and is both environmentally and socially hazardous. The aim of this study was to determine whether the classification of Synthetic Aperture Radar (SAR) imagery could detect and map small-scale mining in Ghana by analyzing multi-temporal filtering applied to three SAR datasets and testing five machine-learning classifiers. Using an object-based image analysis approach, we were successful in classifying water bodies associated with small-scale mining. The multi-temporally filtered Sentinel-1 dataset was the most reliable, with kappa coefficients at 0.65 and 0.82 for the multi-class classification scheme and binary-water classification scheme, respectively. The single-date Sentinel-1 dataset has the highest overall accuracy, at 90.93% for the binary water classification scheme. The KompSAT-5 dataset achieved the lowest accuracy at an overall accuracy of 80.61% and a kappa coefficient of 0.61 for a binary-water classification scheme. The experimental results demonstrated that it is possible to classify water as a proxy to identify illegal mining activities and that SAR is a potentially accurate and reliable solution for the detection of SSM in tropical regions such as Ghana. Therefore, using SAR can assist local governments in regulating small-scale mining activities by providing specific spatial information on the whereabouts of small-scale mining locations.

Cross-platform classification of level and deformed sea ice considering per-class incident angle dependency of backscatter intensity

Abstract. Wide-swath C-band synthetic aperture radar (SAR) has been used for sea ice classification and estimates of sea ice drift and deformation since it first became widely available in the 1990s. Here, we examine the potential to distinguish surface features created by sea ice deformation using ice type classification of SAR data. Also, we investigate the cross-platform transferability between training sets derived from Sentinel-1 Extra Wide (S1 EW) and RADARSAT-2 (RS2) ScanSAR Wide A (SCWA) and fine quad-polarimetric (FQ) data, as the same radiometrically calibrated backscatter coefficients are expected from the two C-band sensors. We use a novel sea ice classification method developed based on Arctic-wide S1 EW training, which considers per-ice-type incident angle (IA) dependency of backscatter intensity. This study focuses on the region near Fram Strait north of Svalbard to utilize expert knowledge of ice conditions during the Norwegian young sea ICE (N-ICE2015) expedition. Manually drawn polygons of different ice types for S1 EW, RS2 SCWA and RS2 FQ data are used to retrain the classifier. Different training sets yield similar classification results and IA slopes, with the exception of leads with calm open water, nilas or newly formed ice (the “leads” class). This is caused by different noise floor configurations of S1 and RS2 data, which interact differently with leads, necessitating dataset-specific retraining for this class. SAR scenes are then classified based on the classifier retrained for each dataset, with the classification scheme altered to separate level from deformed ice to enable direct comparison with independently derived sea ice deformation maps. The comparisons show that the classification of C-band SAR can be used to distinguish areas of ice divergence occupied by leads, young ice and level first-year ice (LFYI). However, it has limited capacity in delineating areas of ice deformation due to ambiguities between ice types with higher backscatter intensities. This study provides reference to future studies seeking cross-platform application of training sets so they are fully utilized, and we expect further development of the classifier and the inclusion of other SAR datasets to enable image-classification-based ice deformation detection using only satellite SAR.

A Gaussian Process Classification and Target Recognition Algorithm for SAR Images

Synthetic aperture Radar (SAR) uses the relative movement of the Radar and the target to pick up echoes of the detected area and image it. In contrast to optical imaging, SAR imaging systems are not affected by weather and time and can detect targets in harsh conditions. Therefore, the SAR image has important application value in military and civilian purposes. This paper introduces the classification of Gaussian process. Gaussian process classification is a probabilistic classification algorithm based on Bass frame. This is a complete probability expression. Based on Gaussian process and SAR data, Gaussian process classification algorithm for SAR images is studied in this paper. In this paper, we introduce the basic principle of Gaussian process, briefly analyze the basic theory of classification and the characteristics of SAR images, provide the evaluation index system of image classification, and give the SAR classification model of Gaussian process. Taking Laplace approximation as an example, several classification algorithms are introduced directly. Based on the two classifications, we propose an indirect multipurpose classification method and a multifunction classification method for two-pair two-Gaussian processes. The SAR image algorithm based on the two categories is relatively simple and achieves certain results.

Snowmelt retrieval algorithm for the Antarctic Peninsula using SAR imageries.

The classification of Synthetic Aperture Radar (SAR) images by knowledge-based algorithms with elevation and backscatter thresholds were used in several studies to detect the Wet Snow Radar Zone (WSZ) in the Antarctic Peninsula. To identify it more accurately based on its seasonal variations, this study proposed the additional use of a threshold in synthetic images, created by rationing summer and winter sigma linear images. In our algorithm we used the following thresholds to detect the WSZ in Envisat ASAR imageries, using the Radarsat Antarctic Map Digital Elevation Model as ancillary data: i) -25 dB < s0 < -14 dB; ii) slinear summer / slinear winter < 0.4; iii) elevation H < 1,200 m for northern tip and H < 800 m for southern tip of the Antarctic Peninsula. The classified images were post-processed by a focal majority 5 x 5 filter and superimposed by an image of rock outcrops derived from the Antarctic Digital Database. The ratio image threshold allowed discriminating the WSZ from the Dry Snow Radar Zone and radar shadows, as well as transitional areas between this glacier zone and the Frozen Percolation Radar Zone, which would be classified incorrectly if we used only elevation and backscatter thresholds.

AGSDNet: Attention and Gradient-Based SAR Denoising Network

Synthetic aperture radar (SAR) images are mainly corrupted by speckle noise, which needs to be removed for further processing. In this letter, we propose an attention and gradient-based SAR denoising network (AGSDNet) to remove speckle noise from SAR images while preserving finer details. In the proposed network, gradient information of the noisy image is first concatenated with its features in order to increase the feature information content (map). An intermediate feature denoising block (FDB) is then employed to reduce noise from this feature map. Finally, two attention blocks, designed and deployed, in the network focus on preserving the more informative features in the image thereby generating a feature preserved denoised image. The proposed network is compared with several classical and deep learning-based SAR denoising methods to demonstrate its superiority in terms of qualitative as well as quantitative measures. We have provided the training and testing code of AGSDNet at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/RTSIR/AGSDNet</uri> .

SPAN: Strong Scattering Point Aware Network for Ship Detection and Classification in Large-Scale SAR Imagery

Ship detection and classification in synthetic aperture radar (SAR) images play a vital role for wide applications. Due to the unique SAR imaging mechanism, ship detection and classification tasks have faced numerous challenges, such as land interference, image defocus, and noise. Many detectors and classifiers have been presented to handle these problems. However, the general deep learning-based detectors and classifiers lack the combination of SAR characteristics, which leads to poor performance. Compared with optical images, SAR images lack the texture information of ships, which brings great difficulties to the recognition task. To address the above issues, a novel deep learning-based ship detection and classification network combined with scattering characteristics is proposed in this article. First, to accurately locate ships in large-scale SAR images, this article designs a strong scattering point aware network (SPAN) by capturing the strong scattering points that existed in the ship area. SPAN recognizes the ship category according to their distribution characteristics. Second, to compensate for the feature loss caused by the down-sampling operation, this article designs a more suitable resolution recovery module to replace the bilinear interpolation method. Third, a region of interest automatic generation module is proposed to fully utilize the axis-align feature of oriented proposal boxes and the sufficient information of horizontal proposal boxes. Furthermore, the classification encoder module extracts the distribution feature of scattering points to classify SAR ships. Finally, the comprehensive experiments in the large-scale dataset for ship detection and classification in SAR images (LDSD) demonstrate the superior performance of the proposed method.

Nonlinear Chirp Scaling Imaging Method for Three-Dimensional Foresight Linear Array Maneuvering SAR

Linear array maneuvering synthetic aperture radar (SAR) can push the limitations of classical SAR on foresight 3-D imaging, which has captured the attentions of worldwide radar researchers for a long time. To reconstruct the position of foresight target, this paper puts forward a nonlinear chirp scaling method for 3-D foresight linear array maneuvering SAR (FLAM-SAR). According to the novel geometry configuration, a nonlinear chirp scaling operation is adopted to equalize the space-variant Doppler frequency modulation rate. The simulation results of the array point scatterers and real airplane target model are used to prove the validity of theoretical derivation and validate the 3-D imaging capacity of FLAM-SAR.

A Low-Power, Ambiguous Synthetic Aperture Radar Concept for Continuous Ship Monitoring

This article addresses the design of a low-cost synthetic aperture radar (SAR) for a single dedicated application, namely illegal vessel detection, which can be implemented using a small satellite and is characterized by reduced transmit power and high resolution. Minimum requirements in terms of noise equivalent sigma zero and resolution that ensure acceptable detection performance are derived on the basis of ship statistics extracted from TerraSAR-X data. One peculiarity of the design is that a pulse repetition frequency much smaller than the nominal Doppler bandwidth is selected to increase the swath width beyond the classical SAR limitation without using digital beamforming, as the azimuth ambiguities can be tolerated for this specific application. Several design examples of SAR systems operating in <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">X</i> -band demonstrate the potential of this concept for small ship monitoring over swaths of 50–90 km with antennas smaller than 0.6 m <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> and very low average transmit powers comprised between 20 and 80 W.

Multi-Class Dual- Stream Convolutional Neural Network for Synthetic Aperture Radar Automatic Target Recognition of Ground Military Vehicle

Synthetic Aperture Radar (SAR) target classification is one of the largest branches of SAR image analysis. Despite the remarkable achievements of deep learning-based SAR target prediction algorithms, current object recognition algorithms are limited in terms of military applications. Acquisition and labeling of SAR target images are time-consuming and cumbersome. Obtaining adequate training data is also challenging in many cases. Deep learning-based models are always susceptible to overfitting because of insufficient training data. This limitation prevents them from being widely used to classify SAR targets. To overcome the problem of insufficient sampling and to learn more accurate representations for SAR image recognition, we propose a two way input of SAR images into a dual stream of DCNN. Concatenating two input SAR images representations was done using the restricted raw SAR data in order to extract the integral features from the 2 input SAR images representations for classification. The proposed methodology addressed the problem of insufficient sample in SAR target classification and improved classification accuracy without overfitting. Experimental results confirmed that the proposed method is effective in addressing the problem of insufficient sample in SAR target classification. This technique can be integrated into any SAR classification model based on convolutional neural networks (CNNs). The model MCDS-CNN results in a 99.1% recognition accuracy. Despite the limited availability of SAR image data from MSTAR, this approach provides good recognition results.

High Resolution SAR Image Classification Using Global-Local Network Structure Based on Vision Transformer and CNN

High-resolution (HR) synthetic aperture radar (SAR) image classification is a challenging task for the limitation of its complex semantic scenes and coherent speckles. Convolutional neural networks (CNNs) have been proven the superior local spatial features representation capability for SAR images. However, it is hard to capture global information of images by convolutions. To solve such issues, this letter proposes an end-to-end network named global&#x2013;local network structure (GLNS) for HR SAR classification. In the GLNS framework, a lightweight CNN and a compact vision transformer (ViT) are designed to learn local and global features, and two types of features are fused in quality to mine complementary information through the fusion net. Then, our research devolves the twofold loss function to reduce the interclass distance of SAR images, which brings more compactness to classification features and less interference of coherent speckles. Experimental results on real HR SAR images indicate that the proposed method has more strong feature extraction capability and noise resistance performance. This method achieves the highest classification accuracy on both datasets compared with other related approaches based on CNN.

Argan Tree (Argania spinosa (L.) Skeels) Mapping Based on Multisensor Fusion of Satellite Imagery in Essaouira Province, Morocco

Tree species identification and their geospatial distribution mapping are crucial for forest monitoring and management. The satellite-based remote sensing time series of Sentinel missions (Sentinel-1 and Sentinel-2) are a perfect tool to map the type, location, and extent of forest cover over large areas at local or global scale. This study is focused on the geospatial mapping of the endemic argan tree (Argania spinosa (L.) Skeels) and the identification of two other tree species (sandarac gum and olive trees) using optical and synthetic aperture radar (SAR) time series. The objective of the present work is to detect the actual state of forest species trees, more specifically the argan tree, in order to be able to study and analyze forest changes (degradation) and make new strategies to protect this endemic tree. The study was conducted over an area located in Essaouira province, Morocco. The support vector machine (SVM) algorithm was used for the classification of the two types of data. We first classified the optical data for tree species identification and mapping. Second, the SAR time series were used to identify the argan tree and distinguish it from other species. Finally, the two types of satellite images were combined to improve and compare the results of classification with those obtained from single-source data. The overall accuracy (OA) of optical classification reached 86.9% with a kappa coefficient of 0.84 and declined strongly to 37.22% (kappa of 0.29) for SAR classification. The fusion of multisensor data (optical and SAR images) reached an OA of 86.51%. A postclassification was performed to improve the results. The classified images were smoothed, and therefore, the quantitative and qualitative results showed an improvement, in particular for optical classification with a highest OA of 89.78% (kappa coefficient of 0.88). The study confirmed the potential of the multitemporal optical data for accurate forest cover mapping and endemic species identification.

SAR Image Classification Using Contrastive Learning and Pseudo-Labels With Limited Data

Synthetic aperture radar (SAR) image classification based on deep learning methods has attracted huge enthusiasm among researchers. Sufficient training data guarantee the generalization ability of neural networks. However, it is quite difficult to collect large amount of high-quality and labeled SAR data in real situations. In this letter, a method based on contrastive learning and pseudo-labels is proposed to solve the few-shot SAR image classification problem. First, a siamese structure is introduced to learn semantic representations of SAR target images. The feature vectors obtained in this stage can reveal the similarity between bi-temporal SAR images. Then, the backbone network of the siamese network is connected to a classification head for instance-level classification. The pseudo-label strategy is adopted to gradually improve the performance. A loss function that changes with iteration is designed to optimize the training process with true labels and pseudo-labels. Besides, considering the imprecision of single-network architecture, we adopt dual-networks and cross-training strategies. The final result is obtained by integrating the results of the dual-networks. Ablation studies show that the designed components of the proposed method are effective. Comparative experiments on the moving and stationary target acquisition and recognition (MSTAR) data set demonstrate the superior performance of the proposed method when dealing with limited training data in the SAR classification.