This work seeks to incorporate recent advances in computer vision, new data acquired through IRWIN (Integrated Reporting of Wildland-Fire Information) and causal ML to get spatially-corrected estimates on recent wildfires for the original ‘stratified-cost-index’, an ex-ante estimate of wildfire expenditure - and understand how private ‘values at risk’ impact resource allocation, controlling for local fire risk factors. For thirty years; under the pressure of rising wildfire suppression costs, the US government has tasked the forest service and other responsible agencies to provide an ex-ante estimation for expected fire costs in order to allow fire managers to approximate their own fire expenditures and determine whether those expenditures fall within historical norms. Models in use currently rely on excluding class A-D wildfires that burn fewer than 300 acres, and feature differential estimates for per-acre suppression costs in the Eastern and Western United States. However, restricting suppression cost estimates to large fires ignores wildfires that have high per-acre costs due to aggressive initial-attack strategies, fires occurring in well-managed forests, and those with hard-to-solve supply chain issues. Using a compact vision transformer to emulate fire-managers’ spatial-information sets, this work increases R-squared of cost forecasts by 25 percentage points, from 65 to 90%, while decreasing correlation with community income, relative to existing methods. By using a regression discontinuity on fires just around the boundary for cost-monitoring eligibility, this work also finds that cost monitoring has a significant and positive impact on the association between per-acre expenditure and property value, implying cost monitoring is a significant factor in explaining why suppression effort appears to benefit primarily wealthy communities.

Soon to come…