A Stacked Stereo Graph Convolutional Network for SAR Image Change Detection

Graph convolutional network (GCN), as a powerful semi-supervised graph learning algorithm, is effective for graph-based feature extraction and classification. It has great potential in extracting and classifying high-ordered spatial structure information in synthetic aperture radar (SAR) images. In this paper, to detect the complicated structural changes in SAR images, a stacked stereo graph convolutional network (SSGCN) is built by combining a stacked stereo graph model (SSGM) with the GCN. The SSGM is built to extract the spatial, temporal and hybrid spatial-temporal information between bi-temporal SAR images. Then the SSGM and GCN are concatenated to construct a hierarchical SSGCN model to execute further the feature depiction and classification. The spectral graph convolution and traditional 2-D convolution are implemented layer by layer to infer the labels of the graph nodes. To show the effectiveness of the SSGCN change detection method, SAR images obtained at the area of Chaohu basin, Anhui, China are tested in the experiments.