A Fine Agricultural Flood Segmentation Model for HJ-2E S-Band SAR Data

Synthetic aperture radar (SAR), with its all-weather, all-time observation capabilities and unique backscattering characteristics for water bodies, has become a primary data source for flood mapping. The Huanjing-2 (HJ-2E), jointly developed by China’s Ministry of Ecology and Environment and Ministry of Emergency Management, provides abundant S-band SAR data, offering strong support for environmental disaster and emergency response. While deep learning methods are now extensively employed for SAR flood segmentation tasks, existing models still face challenges when segmenting discontinuously distributed small flooded areas under agricultural environments, including insufficient classification accuracy, poor structural integrity, and blurred boundaries. To address these issues, this study focuses on high-precision flood mapping in agricultural areas by integrating deep learning methods with HJ-2E S-band SAR data. The first S-band SAR flood segmentation dataset is constructed, derived from HJ-2E data, and includes horizontal transmit-horizontal receive (HH) and horizontal transmit-horizontal receive (HV) dual-polarization backscattering features. Subsequently, a novel deep learning model, S-band flood segmentation U-shaped network (SFSUNet), was designed for flood segmentation using S-band SAR data. SFSUNet adopts the Hiera network from the visual foundation model SAM2 as its backbone and integrates the adapter module to enable efficient cross-domain knowledge transfer from natural images to S-band SAR data. In addition, a full-level channel–spatial interaction block is designed to enhance the modeling and extraction of features for small, fragmented flood areas in agricultural regions. Finally, the effectiveness of the proposed method was validated through experiments in two agricultural flood-prone areas. Experimental results demonstrated that the proposed SFSUNet model significantly outperformed ten existing models. It achieved an intersection over union of 91.87%, outperforming the best comparative method by 3.15%, with precision and recall reaching 96.41% and 95.13%, respectively. Furthermore, experimental analysis of different polarization inputs revealed that the combination of HH and HV dual-polarization data effectively enhanced flood segmentation accuracy. In single-polarization scenarios, HV polarization exhibited superior performance to HH polarization in identifying flooded agricultural parcels. This study provides critical technical support for the application of HJ-2E data in agricultural flood monitoring and lays a foundation for future research and applications of S-band SAR.