CSDBF: Dual-Branch Framework Based on Temporal-Spatial Joint Graph Attention With Complement Strategy for Hyperspectral Image Change Detection

Hyperspectral image (HSI) change detection (CD) aims at obtaining internal components' change information of land cover and land use. In recent years, the development of convolutional neural networks (CNNs) has greatly promoted the research progress in this field. However, the fixed small-size convolution kernels used by CNNs have severely limited the receptive field of information. Another defect of most CNN-based models is their strong dependence on samples, and they are not competent for tasks with a small number of samples. Besides, the traditional CNN-based models can only perform convolution to learn the spatial-spectral features in the Euclidean space, which is not conducive to capturing the geometric changes in land covers in the HSIs. Differently, the graph attention network (GAT) has come into prominence due to its ability to capture the holistic topology structure of images flexibly, and the attention coefficients can be used to effectively model the long-range correlations between land covers. The semi-supervised nature of GAT is also well-suited to handle HSI-CD tasks with limited samples. Nevertheless, the pixel-level topology structure often generates expensive computational costs. To this end, a dual-branch framework based on temporal-spatial joint graph attention (TSJGAT) with complement strategy (CSDBF) is proposed for HSI-CD, which extracts superpixel- and pixel-level features from bitemporal HSIs in parallel and enables them to complement each other. The proposed CSDBF mainly consists of two branches: superpixel-level feature extraction branch (S-branch) and pixel-level feature extraction branch (P-branch). In the S-branch, we introduce the idea of GAT into HSI-CD for the first time and propose a novel TSJGAT module. Thus, the temporal-spatial features of HSIs are propagated and aggregated on the nonlinear graph structure, which makes the changed regions more discriminable. In the P-branch, pixel-level features are obtained by CNNs to correct uncertain factors caused by superpixel segmentation in the S-branch, which is complementary to the S-branch and lays a foundation for more accurate CD. Abundant experiments show that compared with other pioneer methods, the proposed CSDBF can improve the Kappa coefficient by more than 1.9% and 2.5% on average in general sampling rate situations and a low sampling rate situation, respectively, which shows better robustness and better detection accuracy than most existing state-of-the-art methods. The source code of this article can be downloaded from https://github.com/zkylnnu/CSDBF.