Machine learning-based regional flood frequency framework for climate resilient infrastructure

Traditional design flood estimation approaches assume stationary conditions and rely on historical climate data, potentially under- or over-estimating flood risks. This study proposes a regression-based Regional Flood Frequency Analysis (RFFA) framework to assess projected changes in spring flood quantiles under various warming scenarios. The methodology integrates five distinct regression models using Bayesian Model Averaging (BMA) to establish functional relationships between flood quantiles and catchment physio-climatic variables. Focusing on Ontario’s prevalent spring floods, the BMA-RFFA framework demonstrates satisfactory predictive performance across flood quantiles. Future flood quantiles are projected using climatic data from CanRCM4 large ensemble simulations, capturing ensemble spread under a single model and scenario. Results indicate a nearly even distribution of median flood quantile changes from −13 % to +20 % under different warming scenarios, with increases in southern and western regions and decreases in the central region, highlighting spatial variability. Climatic variables such as seasonal water storage and rainfall intensity are identified as key predictors of future flood behavior. Continuous design flood changes over 10-year intervals from 1950 to 2100 were derived using an ensemble pooling approach, capturing the temporal evolution of flood quantiles driven by climate ensemble variability. The study also examined the contributions of Anthropogenic Climate Change (ACC) and Internal Climate Variability (ICV), revealing ICV as dominant in most catchments, while ACC shows stronger influence in western regions. Overall, the proposed framework provides enhanced estimation of future flood quantiles and a robust assessment of their changing behavior under projected climate variability, supporting improved flood risk assessment and infrastructure resilience.