Spatially Explicit Model Predicting Residual Vegetation Patch Existence within Boreal Wildfires

Wildfires are frequent boreal forest disturbances, and particularly in Ontario, emulating them with forest harvesting has emerged as a legislated forest management goal. Since wildfires typically contain a considerable number of unburned residual patches, we present means for learning their characteristics to improve the subsequent emulation of wildfires. We present a method for developing probability maps for the existence of residual vegetation within wildfiredominated landscapes. We use the Random Forests ensemble learning approach to predict the occurrence and distribution of residual patches based on selected predictor variables. Satellitederived data is partitioned into training and validation components using a holdout approach; the model is constructed and calibrated using the training data and evaluated with the validation data. The predictive power of the model is examined using a threshold-independent measure of model performance at five spatial resolutions (4, 8, 16, 32, and 64 m, hereafter described as R4, R8, R16, R32, and R64 respectively). The predictive performance of the model ranges from good (at R64) to excellent (at R4) discrimination ability for one of the largest fire events (F01). The lowest predictive performance is observed for the smallest fire event (F02). Background and Relevance Wildfire in boreal forests is usually intense and frequent, and consumes substantial forest cover but does not burn the entire landscape (Perera, Remmel, Buse, & Ouellete, 2009). Owing to the variation in the geo-environmental factors that affect fire spread, there are areas that partially or entirely escape fire, forming post-fire residual patches. A post-fire residual patch is conceptually defined as a mix of live (and dead) vegetation that forms a spatial continuum, ranging from undisturbed patches of live trees to a single stem (Swystun, Psyllakis, & Brigham, 2001). Understanding the existence and distribution of residual patches involves the need to assess the combined effects of various environmental factors. This lays the framework for assessing the ecological values of residual patches. Spatially explicit information about post-fire forest characteristics is also essential for developing land management policies in forested landscapes. Specifically in Ontario, mapping the characteristics of post-fire residual patches has become a primary requirement for emulating forest disturbances, emerging as a general forest management goal within disturbance driven landscapes (Perera et al., 2009). The development of a framework for real world applications that emulates natural disturbances requires timely and spatially explicit information on residual occurrence. Such maps can be obtained using a predictive modeling approach by merging satellite-based information with ancillary data. Knowing site conditions at which residual patches are likely to occur (or not) forms a basic component of natural resource management and ecological research (Beauvais, Keinath, Hernandez, Master,