TFSANet: Temporal Fusion and Structure-Aware Network for Remote Sensing Change Detection

Remote sensing change detection is important for evaluating disaster impacts, urban monitoring, and land-use analysis. However, existing methods mainly rely on single difference representations. It always lead to constrained discriminative power and pseudo-change artifacts. At the meantime, the class imbalance between changed and unchanged regions further results in optimization bias toward background regions. Considering the above issues, we propose a novel remote sensing change detection network, namely TFSANet. Firstly, a Temporal Feature Fusion Module (TFFM) is designed, which aims to encode model multi-dimensional temporal features through adaptive gating. It can mitigate the restricted feature diversity and spurious responses exists in traditional difference-dominated modeling strategy. Second, a Pixel-Region Balanced (PRB) Loss is proposed to achieve more stable and accurate detection. It optimizes pixel-level classification and region-level structural coherence together, which help suppress background interference and enhance the learning of sparse change regions. Experiments on multiple public datasets show that our method achieves SOTA performance compared with existing methods.