Unsupervised change detection in PolSAR images using siamese encoder-decoder framework based on graph-context attention network

Extracting difference features is a key technique for polarimetric synthetic aperture radar (PolSAR) image change detection. Although the current PolSAR change detection algorithms based on convolutional neural networks (CNNs) can capture the local information of difference features well, the global structure information cannot be extracted effectively, resulting in low detection accuracy. In this paper, we propose a graph-context attention-based siamese encoder-decoder network (GCA-SEDN) for unsupervised change detection in PolSAR images. The GCA-SEDN can mine local and global polarization features simultaneously. Firstly, based on the local features extracted by CNN, the feature optimization graph attention (FOGA) module is constructed to capture global features of PolSAR images. At the same time, the FOGA module greatly refines the image structure feature representation and extracts more discriminative features. Secondly, the designed context -aware dilated pyramid (CADP) module uses multiple dilated group convolutional layers to further extract deep data features with different receptive fields. The obtained multi-scale context data features can adapt well to change targets of different sizes. Finally, by considering both the reconstruction error of the dual-branch encoder-decoder network and the pixel-level classification error, a new hybrid loss function is constructed so that the GCA-SEDN can fully learn change features, thus effectively improving the accuracy of label prediction. Experiments on five real Gaofen-3 PolSAR datasets prove that the proposed GCA-SEDN is more competitive than other existing representative algorithms.