Spatial–Spectral and Channel–Spectral Differences Integration Network for Remote Sensing Image Change Detection

Deep learning methods have shown promising performance in remote sensing (RS) image change detection (CD), largely due to their excellent ability to explore local and global contexts. However, existing methods fail to effectively integrate spectral information into difference features, which makes CD results more susceptible to pseudo-changes. To address the above-mentioned challenges, this article proposes a spatial–spectral and channel–spectral differences integration network (SCDI-Net), which enriches the discriminative representation of change regions. First, the spatial–spectral and channel–spectral differences generation module is designed to explore valuable difference clues from the spatial–channel–spectral perspective for mitigating the effect of pseudo-changes. The deformable feature pyramid network is constructed to aggregate multiscale spatial–spectral and channel–spectral differences. Then, the encoder–decoder is used to model the global context of the difference features. The most critical tokens in spatial–spectral and channel–spectral differences are captured by the top-k token selection and reinforcement module, which eliminates irrelevant tokens and reinforces the feature representation. Finally, the multidirectional attention module facilitates interaction between spatial–spectral and channel–spectral differences, improving the understanding of change regions. Extensive experiments on three datasets show that SCDI-Net effectively detects changes in complex RSCD tasks.