When
Where
Speaker: Ana Fernandez Sirgo, Applied Mathematics GIDP
Title: Machine Learning Prediction of Burn Severity in Southwest US Forests to Support Prefire Assessments of Postfire Debris-Flow Hazards
Abstract: Assessing postfire flow hazards depends on knowing how severely a watershed has burned. Burn severity is used directly in many hydrologic and geomorphic models or indirectly to parameterize the effects of fire on runoff and sediment transport. Empirical models of postfire debris-flow (PFDF) likelihood, for example, rely on burn-severity metrics such as the differenced Normalized Burn Ratio (dNBR), which is derived from remotely sensed data after a fire. This means that PFDF hazard assessments generally cannot be completed until the fire has already occurred. Developing a way to estimate burn severity before a fire could extend these assessments into the prefire period, providing more time for mitigation and risk-reduction planning.
In this study, we develop an ExtraTrees regression model to predict dNBR using fire-intensity metrics simulated with FlamMap, prefire vegetation characteristics, and topography. The model is developed using five fires that burned between 2018 and 2024 in forested areas of Arizona and New Mexico and is independently evaluated using a 2026 fire in western New Mexico. Predicted dNBR maps are compared with observed burn severity and then used as input to a PFDF likelihood model developed for the Southwest USA. We compare PFDF estimates obtained from predicted and observed dNBR and evaluate their sensitivity to the spatial variation in burn severity. These analyses examine whether burn severity can be estimated before a fire and how useful those predictions are for assessing PFDF hazards in advance.