Integrating Machine Learning Models With Probability Distribution Methods for Extreme Flood Risk Assessment

Climate change exacerbates flooding risks, a frequent and devastating natural calamity, particularly in flood-prone regions such as Pakistan. Constraints on computation and inadequate data might cause traditional flood prediction models to suffer. This paper proposes a novel hybrid framework for enhancing flood risk prediction and mapping by integrating statistical approaches with machine learning methods. Using 23 years of precipitation data, we evaluate six statistical distributions, including the Generalized Extreme Value (GEV) distribution, the Gumbel distribution, the normal distribution, the Log Pearson III (LP III) distribution, the Log-Normal distribution, and the Gamma distribution. Three Goodness-of-fit tests, along with the visualization technique, Quantile plots, were used to identify the most suitable distribution methods. The GEV was identified as the best model for predicting extreme events. Furthermore, by demonstrating its resilience in identifying flood-prone locations, with an AUC of 0.96 and PBIAS of −1.2% (training) and −1.4% (validation), the suggested ensemble demonstrated 98% accuracy in training and 98% accuracy in validation. This suggests a minor underestimation of flood risks with low bias and good model generalization. By combining return period estimates and flood susceptibility mapping, the hybrid model provides valuable insights for disaster management and infrastructure design. Not only in Pakistan but also in other flood-prone areas globally, thereby supporting climate adaptation and flood mitigation techniques.