Program in Applied Mathematics Brown Bag Seminar

Abstract:   Debris flows pose a major threat in mountainous regions following wildfire, impacting lives, infrastructure, and water resources. Burned watersheds are more likely to generate larger, more destructive debris flows than unburned basins, thereby contributing significantly to sediment transport and long-term landscape change. Improving post-fire hazard forecasting requires a better understanding of debris-flow initiation and its environmental controls.

This work presents predictive models of post-fire debris-flow likelihood across the Southwest United States. Using a newly developed dataset of 3,144 rainfall events from approximately 200 watersheds in Arizona and New Mexico, supervised classification models were trained with predictors including rainfall intensity, terrain, burn severity, and soil characteristics. The models included logistic regression, linear discriminant analysis, random forest, and XGBoost. Cross-validated threat scores for these models ranged from 0.36 to 0.41. The logistic regression models, which involved peak 15-minute rainfall, mean watershed slope, and burn severity metrics, provide spatially explicit predictions of debris-flow hazards.

Furthermore, this talk outlines future work, focusing on developing machine learning regression models to predict burn severity, measured by dNBR (differenced Normalized Burn Ratio), before wildfires. Algorithms such as Random Forest, XGBoost, CatBoost, and LightGBM will be used to generate spatially explicit pre-fire dNBR maps. These predictions support fire mitigation planning and serve as inputs for debris-flow hazard models, enabling pre-fire assessments of post-fire debris-flow risk at the watershed scale. Preliminary results from this effort are presented.