Accounting for model uncertainties in statistical forecasts of wildfire parameters

Abstract

Gauging the magnitude of model uncertainty and incorporating model uncertainty into predictions is of critical importance when models are used to inform wildfire-related decisions, where ignoring potential risks threaten human health, property, and the environment. Although techniques exist for addressing model uncertainty, these uncertainties are commonly ignored in most analyses. In this dissertation, I will evaluate the effects of model uncertainty on statistical predictions of wildfire activity in multiple contexts and propose techniques to incorporate these uncertainties into predictions. I will determine how uncertainty in the choice of predictive model and climate model influence forecasts of very-large fire activity in the second half of the 21st century, and integrate this uncertainty using a novel Bayesian model averaging approach to produce robust predictions. I find that when these model uncertainties are accounted for, that one may conclude, across the suite of model choices, that the frequency of very-large wildfires should be expected to increase in most regions of the United States if climate changes are not mitigated. The effects of model uncertainty will also be explored in the context of predicting final wildfire size for individual fires that have no yet finished growing. Specifically, I will gauge how the choice of utility function and the inclusion of growth information that is unavailable early in the wildfire’s life alters the predictive ability of statistical models of final fire size and the stability of the model structure. I find that predictions of fire size can drastically change when new utility functions are considered, particularly in models that use growth information. I also find that the covariates used in the best model are sensitive to the choice of utility function, and that no single model is likely to optimally address the preferences of all wildfire-related decisionmakers they are intended to inform. The results of this analysis that (1) the preferred model will often change when new performance measures are considered, and (2) that the preferred model may change over time. I also present a method of integrating the model uncertainties associated with time-varying covariates and ill-defined utility functions into a single predictive distribution using Bayesian model averaging. I find that this novel model averaging approach generally improves predictive performance across a number of performance measures compared to the individual models contained within it. I discuss how the novel methods developed can be applicable to other forecasting applications and how they might allow wildfire professionals make better decisions.