Overcoming data scarcity in flood susceptibility mapping: a hybrid framework coupling high-performance hydrodynamics and machine learning

Abstract Reliable flood susceptibility assessment under current and future climate-change scenarios is critical for regional disaster preparedness but is often limited by the scarcity of historical inundation records and field-surveyed data. To overcome this limitation, this study introduces a hybrid modeling framework that integrates high-performance hydrodynamics with machine learning (ML) to synthesize robust training datasets. Utilizing the GPU-accelerated Two-dimensional Runoff Inundation Toolkit for Operational Needs (TRITON), we simulated 275 maximum precipitation events derived from historical, Shared Socioeconomic Pathways (SSP), and Global Warming Level (GWL) scenarios. These physics-based simulations served as ground truth for training three ML classifiers: Convolutional Neural Networks (CNN), Random Forest (RF), and Support Vector Machine (SVM). Applied to a flood-prone region in central Taiwan, the framework achieved high predictive accuracy (approx. 0.80) across all models, with RF demonstrating superior stability. The assessment reveals that under climate change scenarios, areas classified as “very-high susceptibility” will expand significantly—from 7% in the historical baseline to 15% under GWL 4.0. This study demonstrates that coupling GPU-based physical models with data-driven algorithms effectively overcomes data scarcity, providing a scalable and scientifically rigorous tool for flood risk management in a changing climate.