Assessment of Logistic Regression Model Performance and Physical Controls on January 9, 2018, Debris Flows, Thomas Fire, California

Abstract

ABSTRACT The 2017–2018 Thomas Fire burned 281,893 acres of land in southeastern Santa Barbara County and southwestern Ventura County. An atmospheric river storm impacted the region on January 9, 2018, producing intense rainfall in the western and northern portions of the burned area and triggering numerous post-fire debris flows (PFDFs). The most destructive and deadly flows inundated the town of Montecito, where 23 people died. Debris flow source and inundation mapping data across the fire provide a rare opportunity to assess the interplay between rainfall intensity, watershed characteristics, geologic conditions, and resulting PFDF occurrence. Mapped data are compared to spatially explicit analyses of 857 drainage basins modeled with the U.S. Geological Survey (USGS) logistic regression model (LRM) for PFDF prediction using 15-minute rainfall thresholds at 50 and 90 percent (P50 and P90) probabilities of exceedance. Results indicate that the LRM successfully predicted nearly every PFDF reaching the basin pour point. However, overall model accuracy was lowered by numerous false-positive responses, even where rainfall depths were far above LRM thresholds. Analyses of basins where rainfall was above P50 thresholds reveal a strong correlation between high false-positive responses and basins experiencing rainfall of less than about 150 to 200 percent of USGS thresholds. These false positives occurred in basins with small (0.02–0.05 km2), steep (≥23°) burned areas and in basins underlain by relatively weak geologic units that weather to produce few boulders. Identified relationships provide a basis for refining and improving existing PFDF hazard assessment modeling.