Deep learning based flood segmentation: evaluating EfficientNet and ResNet for UNet, SegNet and DeepLabV3+

Abstract

Rapid and precise flood extent mapping is vital for effective disaster response, as floods cause infrastructure damage, crop losses, casualties, and major economic disruptions. Conventional hydrological models and remote sensing approaches often suffer from high computational demands, coarse spatial resolution, and limited real-time adaptability. This study offers a systematic comparison of three state-of-the-art deep learning segmentation architectures (UNet, SegNet, and DeepLabV3+), each paired with either EfficientNet or ResNet backbones. Our goal is to identify the most efficient and accurate model for flood-affected region delineation. In experiments, UNet with EfficientNet backbone outperformed alternatives, achieving an accuracy of 0.9705, a precision of 0.9609, a recall of 0.9548, a F1-score of 0.9575 and an IoU of 0.9271. These results highlight EfficientNet’s strong feature extraction and lightweight design, making it ideal for near real-time operational deployment. Overall, our findings offer a practical, scalable solution for real-world disaster management, enabling faster damage assessment, optimized resource allocation, and improved flood response.