We used the Maxent spatial distribution modeling tool in combination with BIOCLIM long-term climate normals and derived water-balance metrics (generated for this project) to produce spatially explicit maps of relative wildfire probabilities with 1-km resolution. In this project, we investigate spatial transferability of a suite of California-based fire models to similar climatic regions in Oregon, which has only incomplete, spatially biased data. One point per fire instance was used for training and testing data, and non-fuels areas were masked out during model generation. We present a multi-model average of 4 bootstrapped models, as well as Fire Return Interval (FRI) predictions for one example common vegetation community.
Final Maxent models are based on an average of a suite of 4 models for each region: the product of 2 different variable sets (32 variable ensemble and 15 variable ensemble), and 2 different fire size thresholds in each region (1,000 acre and 5,000 acre), intentionally giving more weight to the 15 variables and the larger 5,000-acre fires, which are involved in all 4 models.
Results/Conclusions
Fire instances are used to generate baseline relative fire probability maps for California and Oregon, which we present here. One data set comprised of Oregon fire points is geographically restricted, in that it completely lacks fires in such places as the southeastern and north-central portion of state. A second data set of fire instances in Oregon consists of polygons spanning eight years, and these appear geographically consistent, although the number of years of data is limited. By incorporating rich California fire history data, we attempted to overcome some of the limitations of the fire data available for Oregon. As new Oregon fire instance data becomes available, we will be able to better evaluate the spatial transferability of California data.
We compared predicted FRI’s within test regions for various common ecosystem types in northern California and Oregon based on 20- and 50-year average burn rates in California, and found that FRI is sensitive to choice of both spatial and temporal extent used to calculate the conversion factor. These sensitivites may be comparable to or larger than model uncertainty. Sensitivity of FRI to choice of temporal extent is demonstrated in the Sierra Mixed Conifer community, where expected median FRI’s differ by 120 years based on choice of burn rate.