Floodwater Region Segmentation from Post-Disaster Aerial Imagery

Abstract

Floodwater region segmentation from post-disaster UAV imagery is important for rapid damage assessment, emergency response, and situational awareness. In this work, we adapt the FloodNet dataset into a dedicated binary floodwater-segmentation benchmark by extracting pixel-level water masks from its original semantic annotations. To further examine model behavior beyond the adapted benchmark, we also include additional recent flooded aerial images collected online and generate corresponding water masks for supplementary evaluation. Using this dataset, we propose a modified U-Net architecture with an added bottleneck transfer layer to refine deep feature representations before decoding. We compare the proposed model with several baseline segmentation models, including encoder-backbone variants based on ResNet-50, DenseNet-121, EfficientNet-B4, and the original U-Net. Experimental results show that the proposed model achieves the strongest overall segmentation performance, reaching 95.92% Precision, 99.90% Recall, 95.91% IoU, and 97.91% DSC. The results demonstrate that the proposed framework provides accurate and robust floodwater delineation in complex post-disaster aerial scenes.