Accelerated and Interpretable Flood Susceptibility Mapping Through Explainable Deep Learning with Hydrological Prior Knowledge

Flooding is one of the most devastating natural disasters worldwide, with increasing frequency due to climate change. Traditional hydrological models require extensive data and computational resources, while machine learning (ML) models struggle to capture spatial dependencies. To address this, we propose a modified U-Net architecture that integrates prior hydrological knowledge of permanent water bodies to improve flood susceptibility mapping in Northumberland County, UK. By embedding domain-specific insights, our model achieves a higher area under the curve (AUC) (0.97) compared to the standard U-Net (0.93), while also reducing training time by converging three times faster. Additionally, we integrate a Grad-CAM module to provide visualisations explaining the areas of attention from the model, enabling interpretation of its decision-making, thus reducing barriers to its practical implementation.