OPT–SAR Flood Change Detection via Adaptive Symmetric Recalibration and Discrepancy-Guided Fusion
Abstract
Flood mapping from predisaster optical and postdisaster synthetic aperture radar (SAR) imagery is important for emergency response, yet severe cross-modal discrepancy often induces pseudo changes in unchanged areas and weakens change evidence in newly inundated regions. To address this issue, we propose SADRNet, an OPT–SAR flood change detection network based on layerwise discrepancy handling. In particular, an adaptive symmetric recalibration (ASR) module is inserted before fusion to regulate cross-modal relations and suppress nuisance discrepancy at shallow-to-mid levels, while semantically condensed residual discrepancy is selectively exploited at the deepest level through discrepancy-guided fusion (DGF) to enhance flood change representation. Experiments on CAU-Flood, SenForFlood, and MT-Wuhan demonstrate that SADRNet achieves competitive and consistent performance across heterogeneous flood and general change detection scenarios.