Abstract: This study addresses the issues of boundary detail loss and insufficient use of dual-temporal change information in SAR images for flood inundation extraction.A Squeeze-and-Excitation weighted ResNet (SE-wResNet)model is proposed,to improve accuracy and robustness in flood detection.The SE-wResNet model is based on ResNet.It uses a dual-temporal,dual-channel input structure to capture changes between pre-and post-flood SAR images.The model incorporates squeeze-and-excitation(SE) and convolutional block attention module(CBAM) mechanisms to enhance flood-sensitive feature extraction in both channel and spatial dimensions.A multi-scale feature fusion decoder is also introduced to improve boundary preservation and semantic discrimination.The model is tested using RadarSat-2 imagery from Dongdian flood detention area in Tianjin.Experiments show that SE-wResNet outperforms models like U-Net and DeepLabV3+ in precision,recall,and overall accuracy.It achieves precision of 0.984 6,recall of 0.988 8,and overall accuracy of 0.997 8.The results show a significant reduction in false positives and missed detections,especially in complex flood scenarios.This demonstrates superior boundary restoration and robustness.The SE-wResNet model provides a reliable solution for automatic flood inundation extraction from SAR images.Its use of dual-temporal information,attention mechanisms,and multi-scale feature fusion enhances detection accuracy.This model is a robust tool for emergency flood monitoring and assessment.

Key words: synthetic aperture radar, flood inundation area extraction, deep learning, attention mechanism, SE-wResNet