Change Detection in SAR Images Based on Improved Non-subsampled Shearlet Transform and Multi-scale Feature Fusion CNN

Traditional methods for change detection in synthetic aperture radar (SAR) images have difficulty in obtaining results from the generated differential image (DI) owing to speckle noise. In recent years, many deep learning (DL)-based methods have emerged because of their outstanding anti-noise and self-learning ability. However, they are limited by the requirement of abundant high-precision labels. Therefore, in this paper, we propose a novel unsupervised method based on improved non-subsampled shearlet transform (NSST) and multi-scale feature fusion convolutional neural network (CNN) for change detection. First, this method improves the traditional NSST algorithm and proposes a novel pseudo-label generator to obtain more pseudo-labels with higher confidence. It is noteworthy that the more accurate the pseudo-labels are, the better the change detection results will be. Second, this method designs a multi-scale feature fusion block in the network to make the feature images contain more complete information and reduces the number of pooling layers to avoid losing feature image details. The main idea of this method is to eliminate the step of generating the DI and directly obtain results from the original images. The theoretical analysis and final results conducted on three real data sets prove its validity. Furthermore, to verify the generality and potential of the propose method, we apply it to the cross-region change detection and compare it with the supervised method, which achieve satisfactory results.