Paddy Rice Inundation Mapping Based on Multi-Resolution Remote-Sensing Imagery and Semantic Segmentation with a Dual-Attention and Edge-Aware Attention Network

Timely and accurate mapping of flood inundation extents in rice-growing regions is essential for flood damage assessment. However, fragmented fields and heterogeneous land cover in complex terrain continue to pose challenges. This study proposes a two-stage mapping framework that integrates multi-resolution remote sensing imagery with deep-learning-based semantic segmentation. First, a dual-attention and edge-aware attention network (DAEA-Net) was developed to derive a high-accuracy paddy rice map from very high-resolution (VHR) imagery. Second, the flood extent in paddy fields was delineated using the prior-histogram marker-controlled watershed (PH-MCW) method on SAR imagery with area-of-interest (AOI) filtering based on terrain slope and on the permanent water bodies and potentially inundated land types of the land-type survey data. On the test set, DAEA-Net achieved the highest OA, IoU, F1-score, and Kappa among the compared semantic segmentation models, including U-Net, HRNet, and DeepLabV3+, with values of 0.9205, 0.8747, 0.9331, and 0.8352, respectively. PH-MCW segmentation with AOI filtering, which removes terrain-induced false alarms while retaining water surfaces, improved flood-water detection, achieving F1-scores of 0.821 and 0.764 on two flood-validation reference datasets (NDRCC and AIR-PolSAR-Seg-2.0), respectively. The proposed framework provides a semi-automated solution for mapping flood inundation over paddy rice fields in complex terrain, in which the prior water proportion is refined by local inspection before application.