LSFC-Net: Lightweight Spatial-Frequency Collaboration Network with sparse interaction for remote sensing change detection

Deep learning-based remote sensing change detection (RSCD) models have shown significant progress, yet they still face two major challenges. First, most methods struggle to effectively cooperate spatial-frequency information and generate complementary dual-domain features, thereby weakening the structural integrity of change objects. Additionally, the abundance of irrelevant semantic information in remote sensing scenes hinders spatiotemporal difference modeling, and the generation of redundant difference information can result in false positives. To this end, this study proposes a Lightweight Spatial-Frequency Collaboration Network (LSFC-Net) with sparse interaction for RSCD. Specifically, a Spatial-Frequency Context Aggregator (SFCA) is developed to decompose the original input into spatial and frequency streams. The former further refines spatial features, while the latter captures frequency discrepancies and texture details, finally integrating the dual-domain information in a complementary manner. Furthermore, a Feature Sparse Interactor (FSI) is designed to highlight genuine changes and suppress the influence of irrelevant semantics. FSI performs cross-temporal interaction on bi-temporal images through a sparse interaction attention mechanism, dynamically selecting the tokens most relevant to genuine changes. Experimental results show that LSFC-Net outperforms most representative RSCD methods in accuracy and efficiency.