Can Machine Learning Predict Post-Fire Vegetation Recovery?

The rate at which vegetation recovers from wildfires exerts strong controls over the availability of ecosystem services, including water provisioning, biodiversity maintenance, and carbon sequestration. Methods to estimate the rate of recovery are valuable for ecosystem managers for prescribed burn strategies and policymakers considering the role of natural and working lands in climate policy goals. Large-scale methods that apply to the wide variety of ecosystems impacted by wildfire are needed to achieve climate policy goals. Here we focus on a remote sensing approach for fires throughout California from 1985-2007 to assess how well machine learning can predict recovery rate following fire disturbance. We measured the change in annual Landsat-derived gross primary productivity (GPP) for over 6000 fire polygons from Cal Fires Fire and Resource Assessment Program (FRAP) along with burn severity, climatology, and topography. We fitted logarithmic growth curves to fire-induced reduction in GPP (dGPP) by calculating the difference between post-fire GPP and a well-matched unburned control GPP. We filtered the fitted curves for positive rates and an R2 greater than 0.2, resulting in approximately 1400 unique fires with most recovering rapidly in under 5 decades. Finally, we trained 25 random forest models with this data and predicted the recovery rate on an out of sample testing set. Preliminary results indicate an average R2 of 0.08, indicating poor model performance on full fire polygons. Subsequent analyses will use finer resolution burn records to better account for heterogeneity in recovery rate, vegetation type, and burn severity within individual fire perimeters. If successful, the resulting models will allow land managers and policy makers to predict immediate and legacy effects of fire occurrence on ecosystem services.