Topology guided and feature decoupled optical-SAR cross-modal knowledge distillation for rapid post-flood water extent mapping

Timely and accurate flood monitoring is crucial for disaster emergency response and agricultural loss assessment. However, optical imagery is often unavailable during flood events. Although synthetic aperture radar (SAR) is not constrained by illumination or weather conditions, its performance is readily affected by speckle noise and complex land-cover scattering, which limit the accuracy of extraction for small and fragmented flooded croplands. To address modality conflict and misleading supervision in conventional optical-SAR knowledge distillation, this study proposes TopoDec-KD, a cross-modal knowledge distillation framework for post-flood water extent mapping. During training, a large network that is driven by optical features serves as the teacher model. By introducing a shallow signed distance field constraint and a deep frequency domain decoupled alignment module, the framework stably transfers purified water semantics and boundary priors from optical data to a lightweight SAR student network, enabling high fidelity segmentation during inference. Experiments on the public S1S2-Water and C2S-MS Floods datasets show that TopoDec-KD achieves IoU scores of 90.15% and 88.52% and F1 scores of 94.82% and 93.91%, respectively. These results represent gains of 6.13% and 6.28% over a baseline trained only with SAR supervision, and they outperform existing state-of-the-art knowledge distillation methods. In a real farmland flood event in Chuzhou, Anhui Province, China, in 2020, the model achieved an overall accuracy of 94.74% based on 114 field samples. Moreover, compared with the more complex teacher model, the lightweight student network delivered inference more than 4.3 times faster with only about a 0.8% decrease in IoU. These results indicate that the proposed method can generate flooded farmland maps with high accuracy, low computational cost, and rapid response using SAR data and auxiliary farmland layers alone when optical data are unavailable.