Hydroclimatic Preconditioning Enables Operational Seasonal Wildfire Forecasting Across Western U.S. Ecoregions

Abstract

Wildfire frequency and severity have increased across western united states ecoregions creating an urgent need for predictive frameworks capable of forecasting seasonal burned area prior to the fire season we developed an end to end machine learning pipeline that integrates winter hydroclimatic observations historical wildfire perimeters and concurrent summer meteorological drivers to predict total ecoregion season burned area across epa level three ecoregions daily snow water equivalent observations from nrcs snotel stations wildfire perimeters from the national interagency fire center and climate variables from gridmet and prism were spatially aggregated using geopandas to construct a regional dataset spanning 2000 to 2022 n equals 460 ecoregion year observations forecast inputs were evaluated as of may 1st to reflect pre season decision constraints structural equation modeling confirmed physical lag pathways between winter snowpack deficits and summer fuel drying while an extreme gradient boosting model trained on antecedent snowpack vapor pressure deficit temperature and wind metrics achieved the primary predictive skill model performance was evaluated using leave one year out cross validation outperforming persistence and linear baseline models shapley additive explanations quantified nonlinear feature interactions driving predictions the integrated framework achieved an out of sample r squared of 0.941 adjusted r squared of 0.939 rmse of 25932687 meters squared or roughly 6408 acres and mae of 20025937 meters squared or roughly 4948 acres high predictive skill was driven by regional spatial pooling and non linear climate coupling rather than data leakage winter snowpack deficits served as a critical preconditioning factor while summer vapor pressure deficit and temperature anomalies amplified burn magnitude the trained pipeline was deployed as an interactive web application generating operational seasonal wildfire risk projections to support proactive land management.