Modeling Forest Wildfire Risks with Non-structural Correction for Spatio-temporal Autocorrelation: A Block Bootstrapping Approach

Our study focuses on modeling wildfire damage in the State of Florida. The approach is to evaluate wildfire risks in a spatio-temporal framework. A block bootstrapping method has been proposed to construct a statistical model accounting for explanatory variables while adjusting for spatial and temporal autocorrelation. Although the bootstrap (Efron 1979) method can handle independent observations well, the strong autocorrelation of wildfire risks brings about a major challenge. Motivated by bootstrapping overlapped blocks methods in an autoregressive time series scenario (Kunsch 1989) and block bootstrapping method of dependent data from a spatial map (Hall 1985), we have developed a method to bootstrap overlapping spatio-temporal blocks. By selecting an appropriate block size, the spatial-temporal correlation can be eliminated. With our saptio-temporal block bootstrapping approach, impacts of environmental factors on SPB outbreaks and implications of pine forest management are assessed. Almost all the explanatory variables, including climate factors, forest ecosystem and socio-economic conditions have been detected to have significant impacts. Consequently, our method offers a way to forecast the future burning risks, given the current influential information of a county.