MAA-SFENet: multi-head adaptive attention and spatial feature enhancement network for remote sensing images semantic change detection

Semantic change detection (SCD) analyzes multi-temporal remote sensing images to identify semantic differences in ground objects. However, existing methods struggle to delineate small-scale changes and preserve fine-grained spatial structures due to insufficient local modeling and background interference. To address these issues, this paper proposes a Multi-head Adaptive Attention and Spatial Feature Enhanced Network (MAA-SFENet) for SCD. Specifically, the Multi-head Adaptive Attention (MAA) module dynamically distributes attention to capture fine local variations, while the Attention-driven Feature Correlation (ADFC) module establishes multi-scale semantic correlations to represent complex change patterns. Moreover, the Spatial Adaptive Feature Modulation (SAFM) module suppresses redundant background and enhances discriminative spatial information. Experiments on the JL1 and SECOND datasets demonstrate that MAA-SFENet outperforms five SCD methods, achieving SeK gains of 5.41% ~29.63% and 2.94% ~28.09%, and OA values of 93.75% and 89.11%, respectively, confirming its superior robustness and accuracy.