Flood Area Segmentation: A Deep Learning Framework for Pixel-Level Flood Detection from Aerial Imagery

Abstract

Quickly and accurately identifying flooded land from aerial imagery can assist with disaster response, damage assessment, and long-term planning. This paper examines binary semantic segmentation for flood detection, where each pixel is classified as either flooded or non-flooded. Two versions of the project were developed: one in Python using PyTorch and torchvision, and another in R using R torch. The Python implementation fine-tunes a DeepLabV3 model with a ResNet-50 backbone pretrained on ImageNet. The R implementation uses a custom U-Net trained from scratch, since the R torch workflow does not provide the same readily available pretrained vision models. Both models use the same combined Binary Cross-Entropy and soft Dice loss function, and both are evaluated on the same held-out images. The Python DeepLabV3-ResNet50 model achieves a Dice coefficient of 0.8765 and an Intersection over Union (IoU) of 0.7865, while the R U-Net achieves a Dice coefficient of 0.8003 and an IoU of 0.6889. These results demonstrate that pretrained models are highly effective for flood segmentation, while also showing that a carefully designed U-Net in R can still produce strong performance.