A New Method for Nowcasting Wildfire Risk

Wildfire risk prediction relies on the often-heuristic assessment of diverse fire potential indices, fuel maps, fire weather indices and prior fire activity data. Here we present a model nowcasting daily wildfire genesis probability and expected wildfire sizes in the contiguous US.

Predictors were selected and developed to account for climate, vegetation, topographic and human effects on wildfire genesis. Climate factors are represented by multiple fuel wetting and drying processes at daily to seasonal-scale antecedences, snowpack, and wind. We use GPP to predict fuel mass and recent growth, and dominant vegetation type. Human factors include population, landscape accessibility and ignition sources such as powerlines.

The first stage of the model predicts wildfire genesis probability as a zero-inflated process with an explicit probability of fire preclusion, whilst the second stage models fire sizes according to a generalised extreme value distribution. Nonlinear effects are accounted for via global optimisation for the domain for which each variable drives changes in fire genesis behaviour and the appropriate variable transform.

The model has good predictive and explanatory power, as shown by various performance metrics and the meaningful nonlinear relationships identified in the optimisation process. We show that this method can resolve seasonal wildfire risk dynamics well over smaller ecoregions than the observational record permits, allowing us to quantify the extent to which fire risk is determined by seasonal-scale versus daily-scale effects.