Application of a deep learning framework integrating SAR imagery and terrain slope to flood mapping

Deep learning and remote sensing are critical to flood mapping, with synthetic aperture radar (SAR) offering a weather-independent data source. To improve segmentation accuracy, we designed a GSConv Block and introduced LSK Attention for low-level feature extraction, forming a lightweight network, GLNet. Shadow-induced misclassification was mitigated through feature-level fusion of SAR and terrain slope, leading to the proposed SSFNet. Supplementary samples were generated from land cover products to enlarge the training dataset. Results showed that GLNet achieved 88.57% IoU on the S1-Water dataset, outperforming SegFormer by 1.8%. SSFNet achieved 93.28% IoU, outperforming pixel-level fusion by 3.16%. After expanding the training set, SSFNet achieved R2 > 0.95 and reduced RMSE by 1.4 km2 across 256 sites, demonstrating strong generalization for Chaohu Lake. Applied to the August 2024 flood in Liaoning, it revealed a strong correlation between rainfall and inundation. This study provides support for rapid flood mapping using SAR imagery.