Cross-Level Frequency-Domain Boundary-Aware Lightweight Network for SAR Flood Detection

Synthetic aperture radar (SAR) has become a crucial observation tool for flood disaster monitoring due to its all-weather and all-time imaging capabilities. However, existing bitemporal SAR flood detection methods primarily rely on spatial-domain feature fusion, which often results in missed detections and false alarms when encountering irregular-boundary floods. To address this limitation, this study introduces a cross-level semantic-guided frequency boundary-aware (CSFB) module and a multiscale semantic-aware bottleneck (MSAB) module. Specifically, the CSFB module introduces deep semantic features and analyzes the frequency-domain correlation between shallow features and deep semantic features, thereby adaptively selecting frequency components from the shallow features and fusing them with the deep semantic features. Thus, this design preserves the main structure of the flood body from the shallow features, while enhancing the model's ability to perceive flood boundary details within the high-frequency components. Furthermore, the MSAB module integrates convolutions with various dilation rates into multiscale feature fusion to establish associations between local flood details and global context information, enhancing the unified detection of both large-scale and small-scale flood regions. By integrating CSFB and MSAB modules, a novel cross-level frequency-domain boundary-aware lightweight network (CFB-Net) is proposed to enhance the perception of irregular-boundary floods. Comprehensive experiments on the S1GFloods and ETCI-2021 datasets, along with generalization evaluations on the Jubba and Weihui real-world flood events, demonstrate that CFB-Net effectively enhances the perception of irregular flood boundaries and suppresses false alarms caused by weak-scattering interference.