Distribution-informed machine learning for flash flood susceptibility: integrating weibull extreme value theory with interpretable models.

Flash floods represent one of the deadliest weather-related hazards globally, yet their prediction remains fundamentally challenged by extreme class imbalance in observational data. This study addresses a critical methodological gap: traditional evaluation metrics, both overall accuracy and Area Under the ROC Curve (AUC), are systematically misleading for rare event prediction. We demonstrate empirically how models achieving 93% accuracy and AUC exceeding 0.98 can simultaneously fail to detect 65% of flood events. Moving beyond conventional approaches, we introduce distribution theory-informed feature generation by integrating Extreme Value Theory through Weibull distribution analysis. We derive 24 features from rigorous statistical characterization of precipitation extremes spanning 16 years (2010-2026) of ERA5-Land reanalysis over Nova Scotia, Canada. Evaluating seven model configurations using Environment and Climate Change Canada operational warning thresholds, we find that adding just six Weibull-derived features to a Random Forest baseline nearly doubles flood detection, with recall increasing from 0.35 to 0.65 and F1-score from 0.48 to 0.74, while maintaining 87% precision. This controlled comparison provides the clearest evidence for the value of distribution-informed features. Across architectures, Support Vector Machines with selected features achieve 93.4% balanced accuracy with perfect recall, while Artificial Neural Networks achieve a balanced operational profile (75% recall, 65% precision). SHAP analysis reveals that physically meaningful interaction features, particularly the intensity-duration product and rain-on-saturated-soil, dominate predictions, with raw precipitation ranking only sixth, confirming that models learn genuine multivariate susceptibility structure rather than recovering classification thresholds. These findings provide essential guidance for practitioners: comprehensive reporting of balanced accuracy, precision, and recall is mandatory for imbalanced datasets where traditional metrics mask operational failure.