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

Area of Science:

• Hydrology and Meteorology
• Data Science and Machine Learning
• Extreme Weather Event Prediction

Background:

• Flash floods are a major global hazard, but predicting them is hindered by imbalanced observational data.
• Traditional evaluation metrics like accuracy and Area Under the ROC Curve (AUC) are unreliable for rare events, masking model failures.
• Existing methods fail to adequately capture the complexities of extreme precipitation events.

Purpose of the Study:

• To address the critical methodological gap in evaluating rare event prediction models.
• To introduce and validate distribution theory-informed features for enhanced flash flood prediction.
• To demonstrate the superiority of new evaluation metrics and feature engineering over traditional approaches.

Main Methods:

• Integrated Extreme Value Theory (EVT) using Weibull distribution analysis to generate 24 novel features from 16 years of ERA5-Land reanalysis data for Nova Scotia.
• Evaluated seven model configurations, including Random Forest, Support Vector Machines, and Artificial Neural Networks, against Environment and Climate Change Canada operational warning thresholds.
• Employed SHAP (SHapley Additive exPlanations) analysis to interpret feature importance and model behavior.

Main Results:

• Models incorporating six Weibull-derived features nearly doubled flood detection recall (0.35 to 0.65) and improved F1-score (0.48 to 0.74) while maintaining high precision (87%).
• Support Vector Machines achieved 93.4% balanced accuracy with perfect recall; Artificial Neural Networks showed a balanced operational profile (75% recall, 65% precision).
• SHAP analysis confirmed that physically meaningful interaction features (e.g., intensity-duration, rain-on-saturated-soil) are more critical than raw precipitation for accurate predictions.

Conclusions:

• Distribution theory-informed feature generation significantly enhances flash flood prediction performance, particularly for rare events.
• Comprehensive reporting of balanced accuracy, precision, and recall is essential for imbalanced datasets to avoid masking operational failures.
• The study provides crucial guidance for practitioners, emphasizing the need for advanced feature engineering and appropriate evaluation metrics in extreme weather prediction.