An Intelligent Deep Learning Framework for Flood Risk Mapping from Aerial Imagery

Abstract

Flood detection and damage assessment from aerial imagery play a critical role in effective disaster response and management. However, accurately identifying flooded roads and damaged areas from large-scale aerial and satellite imagery remains challenging due to variations in illumination, occlusions, water reflections, and complex environmental conditions. To address these challenges, a U-Net-based segmentation model was developed to accurately delineate road regions and distinguish between flooded and non-flooded road segments in aerial imagery. The proposed model employs an encoder–decoder architecture with skip connections to retain fine-grained spatial information while effectively learning high-level semantic features, thereby enabling robust segmentation under diverse environmental conditions.