Comparative Analysis of Machine Learning Surrogates for Accurate and Real-Time Flood Prediction

Sammanfattning

The escalating threat of climate-driven flooding (Core Writing Team et al., 2023) creates an urgent imperative for disaster preparedness systems that are both rapid and precise. Traditional hydrodynamic models, such as HEC-RAS, provide the necessary physical fidelity for mapping inundation but are computationally prohibitive for real-time Early Warning Systems (EWS). Conversely, standard library interpolation methods often fail to capture the non-linear complexities of modern flood dynamics, particularly when accounting for multiple hydrologicaldrivers. This Master’s thesis addresses this critical operational gap by developing and rigorously evaluating Machine Learning (ML) surrogate models capable of predicting flood extents instantly without sacrificing hydraulic accuracy. The study focuses on a high-risk 98 km reach of the Po River (Cremona-Borgoforte), utilizing high-fidelity HEC-RAS simulations to generate a synthetic training dataset. To identify the optimal architecture, five distinct regression models were implemented: Multiple Linear Regression, Support Vector Regression, Random Forest, Extreme Gradient Boosting (XGBoost), and Gaussian Process Regression (GPR). These models were subjected to a hierarchical evaluation, testing their performance not only on interpolation within the training range but also on their "Extrapolation Robustness" against extreme, unobserved flood events (450-year and 500-year return periods).T he comparative analysis reveals a critical trade-off between local precision and global robustness. While tree-based ensembles (Random Forest and XGBoost) demonstrated superior accuracy in the interpolation regime, they exhibited a structural "saturation" bias during extreme events, systematically underestimating flood depths for discharges exceeding the training domain. In contrast, Gaussian Process Regression emerged as the robustest architecture. It maintained physical consistency during extrapolation and provided essential uncertainty quantification, offering a spatial map of model confidence. Consequently, this research establishes GPR as the viable candidate for operational EWS, enabling civil protection agencies to generate real-time, risk-informed inundation maps.