Statistical Challenges in Forecasting Wildland Fire Regimes

Wildfires in northern Canada have been very conspicuous in recent years, threatening both valuable infrastructure and human lives, such that small cities with tens of thousands of inhabitants have had to be evacuated on several occasions. As human settlements encroach into the forest, their exposure to wildfire will increase. At the same time, growing industrial activity in the midst of the forest, especially by the energy and forestry sectors, is also at risk; the size of the potential recurring losses are nationally significant. The degree of risk varies considerably from one region to another, but will grow worse in many places under climate warming. In order to insure against or mitigate future losses, it's necessary to quantify and map these risks, which involves forecasting the probabilities of future events. These forecasts must be sensitive to climatic factors (both direct and indirect), and to changes in landcover that result from intentional clearings, forestry, road building, and from fires themselves. In this talk, we discuss first how simple spatial simulation models may be used to generate such forecasts, by means of Monte Carlo simulations that generate repeated samples from predictive statistical models of fire occurrence and size. These statistical models can be estimated from the records of fire management agencies using maps of landcover and interpolated fire weather covariates. Although such existing models are valuable for ecological studies and policy analyses, they probably fall short of the forecast reliability necessary for risk mapping for actuarial purposes. The reasons for this begin with the historical record, which is relatively short, in many places beginning only in 1960s. This record is the product of the fire management agencies. This has several consequences that complicate statistical inference. Firstly, the efficiency of the organised fire detection system has not been constant over the period of record. This must have resulted in imperfect detection, especially of smaller fires, leading to size-biased sampling. Increased detection efficiency in recent decades is associated with increased fire suppression effort. To the extent that these efforts are successful, fires are smaller than they would have been. Both these factors bias statistical models of the relationship between fire activity and climate…the very models that we need for forecasting. We conclude by outlining some efforts underway to correct these biases by applying modern methods of distance sampling, hierarchical modelling and survival analysis.