Automatic Detection and Semantic Segmentation of Flood Events Using Deep Learning-Based Methods

Abstract

Floods rank among the most destructive natural disasters, often resulting in severe human and economic losses. Timely detection and swift intervention are essential to reducing their devastating effects. This study proposes a deep learning-based framework to assess flood impacts by applying classification and semantic segmentation techniques to high-resolution UAV imagery acquired in the aftermath of Hurricane Harvey. Transfer learning-based architectures including MobileNetv2, Xception, Inceptionv3, and DenseNet121 were employed for flood presence classification, while U-Net, UNet++, ENet and Deeplabv3+ models were utilized to segment flooded buildings, roads, and other objects. The results demonstrate that DenseNet121 and Deeplabv3+ architectures achieved high performance in classification and segmentation operations. This study highlights the effectiveness of deep learning-based automated systems in accelerating post-disaster response processes and optimizing human intervention. © 2025 IEEE.