Predictive Models for Postfire Debris Flow Initiation in the Southwest USA

Abstract Postfire debris flows pose a threat to life and infrastructure and significantly contribute to sediment supply in upland channels, thereby impacting water quality, stream habitats, and landscape evolution. Models designed to assess postfire debris‐flow likelihood at the watershed scale in response to design or forecast rainstorms are beneficial for identifying and mitigating postfire debris‐flow hazards, and for modeling the long‐term evolution of steep, fire‐prone landscapes. This study used four machine learning (ML) algorithms, logistic regression, linear discriminant analysis, random forest, and XGBoost, to develop classification models for postfire debris‐flow likelihood at the watershed scale. We compile a new data set of postfire debris‐flow observations from the Southwest USA, specifically from Arizona and New Mexico. The data set includes information related to hydrogeomorphic response (i.e., debris flow or no debris flow) for 3,107 rainfall events that occurred within the first year following fire across 200 watersheds. We use these data to develop two logistic regression models, based on rainfall, terrain, and fire severity metrics, that can provide spatially explicit predictions of debris‐flow likelihood across burned landscapes. Both models use two features that combine peak 15‐min rainfall accumulation with mean watershed slope and a fire severity metric. All four ML algorithms produced models with threat scores ranging from 0.37 to 0.41 when trained using these features, with logistic regression achieving the highest threat score. Results improve our ability to assess postfire debris‐flow hazards in the Southwest USA and provide insight into how rainfall, terrain, and fire severity influence postfire debris‐flow initiation.