Efficient Water Body Detection Based on Knowledge Distillation for SAR Imagery

Abstract

Synthetic aperture radar (SAR) is widely used for water body detection due to its efficiency and ability to operate in all weather conditions. However, its scattering properties and single-polarization limitations pose challenges for data extraction and reduce the accuracy of water body detection algorithms. To mitigate this limitation, recent studies have focused on transforming SAR datasets into electro-optical (EO) image modalities through cross-modal translation models, aiming to enhance multispectral feature interpretability. However, such transformation frameworks require substantial computational power, which compromises the real-time processing capabilities critical for rapid disaster response, such as a flood. In this letter, we propose a lightweight SAR water body detection framework that integrates knowledge distillation and channel attention. A teacher network trained on rich EO data guides an SAR-specific student model, with both employing attention branches. The student’s attention is supervised by the teacher to enhance SAR feature extraction via attention-aligned distillation. Evaluated on the Sen1Floods11 benchmark dataset, our experimental results outperform the baseline model by 3.5% in intersection over union (IoU).