SIFANet: Spatial-Temporal Interaction and Frequency Adaptive Awareness Network for Change Detection in Remote Sensing Images

Change detection (CD) is an essential mission in the realm of remote sensing. In previous years, deep learning has been introduced into the domain of CD and has made great progress. How to effectively utilize useful information to improve detection performance remains a challenge. To alleviate this concern, we propose a network based on spatial-temporal interaction and frequency adaptive awareness. The network contains three main modules. Specifically, we design a spatial-temporal interaction module that enhances the interaction of disparity features with diachronic features to intensify the focus on change regions. Subsequently, in the decoding phase, we use deep features to guide the shallow feature generation, which can effectively filter the background clutter of shallow features, where an adaptive upsampling module is implemented for effective feature fusion. Finally, frequency adaptive awareness module is utilized for modeling multiscale features by combining frequency domain and temporal domain features, thus enhancing the model's ability to perceive changed regions. We have performed experiments over three prevalent datasets CDD, SYSU-CD, and LEVIR-CD, respectively. The proposed method achieves IoU of 95.70% (4.92% improvement over secondary one) on the CDD dataset, 84.34% (1.94% improvement over secondary one) with LEVIR-CD dataset, and 69.89% (0.22% improvement over secondary one) for SYSU-CD dataset. Our approach outperforms other state-of-the-art CD methods. Visible results indicate that our method generates more complete and clearer details of the changes.