Burn-Scar Half-Life

Image Credit: Mike Eliason

On 9 January 2018, a month after the Thomas Fire burned the hills above Montecito, California, a short, intense storm sent debris flows through the town that killed 23 people and closed Highway 101 for almost two weeks. The USGS publishes debris-flow hazard maps for burned areas, and in 2026 showed how that hazard falls as vegetation grows back; what nobody has mapped is how long each road and water intake downstream stays at risk. We will use California fires since 2015, track vegetation recovery in each burned watershed from satellite imagery, and train a model that predicts how many years a watershed takes to recover from its burn severity, slope aspect, elevation and the rain that fell after the fire. The output is a map of years-to-safe for the roads and water intakes below each burn scar, checked against the USGS record of which storms actually produced debris flows.

Research questions

1. For California watersheds burned since 2015, can the number of years until vegetation returns to its pre-fire level be predicted from burn severity, aspect, elevation and post-fire rainfall, and how wide are the uncertainty ranges?

2. When that recovery prediction is folded into the USGS debris-flow likelihood model, how many years does each road segment and water intake below a burn scar stay above the hazard threshold, and did debris flows actually occur in the years the model still called hazardous?

Datasets

• MTBS burn severity (dNBR) for California fires from 2015 onward

• Landsat and Harmonized Landsat Sentinel-2 time series of NBR and NDVI, 2013 onward, for the recovery curves of each watershed

• IMERG rainfall for the months after each fire, as totals and peak intensities

• USGS runoff-generated postfire debris-flow inventory: 8,981 storm-by-basin observations from 80 fires, 1,140 of them debris flows, with rainfall and terrain already attached

• NHD watershed boundaries, OpenStreetMap and Caltrans roads, and surface-water intake locations from the California State Water Resources Control Board, with partner access where the intake data is restricted

Machine learning

• A gradient boosting model that predicts years-to-recovery for each burned watershed (the year NBR returns to within 10 percent of its pre-fire level) from burn severity, aspect, elevation, land cover and first-year rainfall, with quantile regression forest ranges

• The published USGS M1 likelihood equation, re-run each year with the recovery-adjusted burn severity, to turn the recovery curve into a yearly debris-flow likelihood for each watershed

• The M1 equation with no recovery adjustment, as the baseline

GIS

• Delineate the burned watersheds on NHD boundaries and summarize severity, slope, aspect and elevation for each

• Build each watershed's NBR recovery curve from the satellite time series

• Join each watershed's yearly likelihood to the road segments and water intakes downstream of its outlet

• Hold out whole fires when testing, so a watershed is never predicted from its neighbors' data

• Map years-to-safe by county for the public version

References

• Graber, Thomas, Kean, King and Kostelnik (2026). Characterizing changes in postfire debris-flow hazard as burned areas recover. Geosphere.

• Staley, Negri, Kean, Laber, Tillery and Youberg (2017). Prediction of spatially explicit rainfall intensity–duration thresholds for post-fire debris-flow generation in the western United States. Geomorphology 278, 149–162.

• Hoch, McGuire, Youberg and Rengers (2021). Hydrogeomorphic recovery and temporal changes in rainfall thresholds for debris flows following wildfire. Journal of Geophysical Research: Earth Surface 126(12).