Flood Hazard Modelling Assessment Using Deep Learning and Earth Observation

Abstract

This study explores the integration of satellite imagery with deep learning models, specifically the U-Net architecture, for flood hazard modelling and detection in flood-prone regions of Uttar Pradesh, India. The research focuses on using Sentinel-1 satellite data, combined with multispectral imagery, to develop a robust flood detection system. The methodology involves preprocessing satellite images, creating flood polygons through change detection, and training a U-Net model with a ResNet34 backbone on a dataset of 28,896 image patches. The model achieved an accuracy of 92.82% and an Area Under the Curve (AUC) score of 90%, demonstrating strong performance in identifying flooded areas. The study highlights the effectiveness of combining remote sensing and deep learning techniques for flood mapping, providing high-resolution flood maps validated against ground truth data. The results underscore the potential of AI-driven approaches in enhancing flood risk assessment, disaster response, and management strategies. In order to address class imbalance, optimizing threshold selection is necessary and incorporating additional data sources, such as LiDAR and hydrological models, can improve model performance and applicability. This research contributes to advancing flood monitoring and mitigation efforts, offering a scalable and efficient solution for flood-prone regions.