Post wildfire debris flow assessment

Satellite-Driven Machine Learning Pipeline for Sub-Basin Scale Hazard Assessment

Wildfires dramatically increase the risk of post-fire debris flows by stripping vegetation, altering soil hydrology, and reducing infiltration capacity. This project develops a Random Forest-based machine learning pipeline to predict debris flow occurrence at the sub-basin scale following wildfire events in steep, rugged terrain.

The pipeline integrates 37 environmental features derived from satellite remote sensing (Landsat-8/Google Earth Engine), terrain analysis (WhiteboxTools), storm and antecedent rainfall (ERA5-Land, NOAA Atlas 14), and soil properties (SoilGrids v2.0). The model was calibrated and validated against field-verified debris flow inventories from the 2020 Dolan Fire (Big Sur, California), compiled by Cavagnaro et al. (2025).

The approach complements the operational USGS logistic regression baseline (Staley et al., 2017) and achieves an AUC-ROC of 0.906 on held-out data. An operational probability threshold is set below 0.5 to prioritize recall, minimizing the risk of missed debris-flow events in life-safety contexts.

A web-based interactive dashboard is under development to deliver real-time hazard predictions to emergency managers and land-use planners following new wildfire events.

Key Features:

• 37-feature matrix spanning terrain, burn severity, rainfall, and soil properties

• Fire-specific dNBR burn severity calibration

• Sub-basin scale predictions mapped as interactive GeoJSON overlays

• Benchmarked against the USGS Staley et al. (2017) M1 logistic regression model

Team: Subash Poudel, Dr. Rocky Talchabhadel (JSU), Dr. Nawa Raj Pradhan (ERDC/CHL)

Support: ORISE Fellowship, U.S. Army Engineer Research and Development Center (ERDC), Woolpert Digital Innovations, Taylor Engineering, Inc.