STDFNet: A Symmetric Temporal Difference and Multi-Scale Context Gating Network for Remote Sensing Change Detection

Abstract

High-resolution remote sensing change detection must distinguish genuine changes from pseudo-changes caused by misregistration, illumination, and texture variations. We propose STDFNet, a lightweight Symmetric Temporal Difference and Multi-Scale Context Gating Network. A shared MobileNetV2 extracts four-level bi-temporal features. At each level, symmetric temporal difference encoding combines forward, reverse, and absolute differences using a shared encoder and exchangeinvariant sum and gap terms. Multi-scale context gating integrates local and dilated depthwise convolutions with channel-spatial gates, while a deeply supervised top-down decoder reconstructs the change map. With only 2.612M parameters, STDFNet achieves 0.7186 IoU and 0.8362 F1 on SYSU-CD, and 0.6241 IoU and 0.7685 F1 on UAV-CD. Compared with a same-framework absolutedifference baseline, it improves IoU by 2.25 and 3.05 percentage points, respectively. Ablation results confirm the complementarity of the two modules.