Entropy-Guided Adaptive Window Attention for Lightweight Remote Sensing Change Detection

Deep learning models for bi-temporal remote sensing change detection increasingly rely on attention-based context modeling, yet most designs apply a fixed computation pattern to all regions. This uniform processing is inefficient because large homogeneous areas typically dominate high-resolution imagery, while only a small fraction contains structurally complex changes. This letter presents an Entropy-Guided Adaptive Window Attention network (EWACD), a lightweight framework that allocates computation according to local change complexity. The core component is a Dynamic Entropy-Guided Hybrid (DEGH) block, which estimates a change-entropy prior from bi-temporal feature differences and uses a differentiable router to select among three branches with distinct receptive fields: a low-cost depthwise convolution branch for low-entropy regions and two window-attention branches for medium- and high-entropy regions. In this way, EWACD concentrates global context modeling on ambiguous areas while preserving high throughput on easy backgrounds. Extensive experiments on LEVIR-CD and WHU-CD demonstrate a strong accuracy–efficiency trade-off. EWACD achieves 87.20% F1 on LEVIR-CD and 90.02% F1 on WHU-CD, with only 0.28 GFLOPs and 156 FPS. The results indicate its practicality for real-time edge-oriented deployment.