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.

Existing deep learning-based change detection methods in the field of polarimetric synthetic aperture radar (POLSAR) usually directly deal with intensity images. Methods can be easily transferred from optical image processing to synthetic aperture radar (SAR) image processing. However, the polarization information, which is critical in POLSAR data, has been discarded under normal situations. This paper introduces a novel joint change detection network based on similarity learning for coregistered complex-valued SAR data. A pseudo-siamese network takes both amplitude information and polarization information of POLSAR data as the input. The fusion of low-level, middle-level and high-level features enables the network to keep high-resolution and have strong representation ability during training procedures. Our novel sub-networks, which we term C2-Net and Intensity-Net, deal with the covariance matrix of complex SAR data and amplitude SAR data, respectively. The Intensity-Net works as a typical classification network and detects targets directly. The C2-Net attempts to find the relationship between two SAR data patches. An improved cosine similarity function is used to measure the similarity between two generated feature vectors in C2-Net. Output probability vectors of the two sub-networks are combined for final change detection. The two sub-networks are trained jointly and simultaneously. Control experiments show that proposed improvements are working. Experimental results on dual-polarization complex-valued SAR data from Sentinel-1 demonstrate the effectiveness of the proposed method.

Change detection, as an essential part of synthetic aperture radar (SAR) automatic target recognition (ATR) systems, remains a challenging problem. The lack of labeled data seriously impedes the development of deep learning-based methods in SAR applications. In this letter, we propose a patch-based semisupervised method to detect changed pixels from limited training data in the field of SAR. The complete approach includes the unsupervised pretraining stage and the iterative discrimination stage. The states of pixels are determined by comparing the corresponding image patches. First, to reduce the impact of insufficient data, we train a trapezium U-Net-like structure to extract representative and generalized features by unsupervised training on all pixels. Feature maps contain detailed and semantic information due to repetitive feature fusions. A designed feature activation module is utilized to recalibrate fused features. Finally, the discrimination is finished by a simple classification convolutional neural network (CNN) following a two-step training strategy. Ablation studies indicate the function provided by the proposed modifications. Experiments on real SAR images demonstrate that the proposed method can improve the detection accuracy by more than 1.2% compared with other state-of-the-art deep learning-based methods.