SSTGCNs: Spectral-Spatial-Temporal Long-Range Dependencies Joint Feature Extraction With Graph Convolutional Networks for Adaptive Change Detection

Hyperspectral image change detection plays an important role in ground observation tasks, making full use of the rich spectral and spatial information in the bitemporal hyperspectral to identify subtle changes in the surface. Currently, most methods are extracting spatial -spectral features, ignoring the interaction between bitemporal images. In addition, there is a long-distance dependence between the spectrum and pixels of hyperspectral images. How to capture the dependence between long-distance bands and pixels is a problem that needs to be solved at present. To solve the above problems, this paper proposes SSTGCNs: spectral-spatial-temporal long-range dependencies joint feature extraction with graph convolutional networks. First, the network uses Spectral-Spatial-Temporal Feature Extraction Module (SSTFEM) to capture spectral-spatial-temporal features and increase the interactive features of bi-time phase images. Secondly, a long-distance dependency capture module is designed to capture the long-distance dependence between bands and pixels through the Long-Range Spectral-Spatial Dependency Module (LRSSDM). Finally, through Redundancy Suppression and Adaptive Feature Fusion (RSAF), redundant features are removed and high-information features are fused to improve the model's feature expression ability and generalization ability. To verify the effectiveness of the model, we conducted experiments on three public hyperspectral CD datasets. The experimental results showed that the OA and Kappa indicators of our model on the USA data were improved by 0.55% and 1.44% respectively compared with the latest methods. On the China dataset, the proposed method compared the latest AIWSEN, OA and Kappa to obtain suboptimal results, while the P, R and F1 scores achieved the optimal results, which was 2.59%, 0.89% and 1.14% higher than the state-of-the-art method respectively. The OA, P, R, F1 indicators on the River dataset achieve optimal results: 0.36%, 1.32%, 0.5% and 0.91% improvements respectively over the state-of-the-art methods. Explain that our change detection method works better for spectral data containing more information.