Spatially Explicit Prediction of Residual Vegetation Patch Occurrence within Boreal Wildfires

Abstract

Even after landscapes are disturbed by boreal wildfires, considerable quantities of residual vegetation remain. A method for developing spatially explicit predictive probability maps is presented to identify the presence of residual vegetation within burned boreal landscapes in North-western Ontario, where we learned residual presence expectations from a suite of wildfires that burned from 2002-2003. This approach relies on easily measured variables due to the lack of detailed local information in this remote region. We cross-validate predictions within our training data, a suite of 11 wildfires, using a bootstrapping approach (internal validation) and then test the model on an independent event that burned in 2011 (external validation). The predictive model is based on the Random Forest algorithm and is implemented at 5 separate spatial resolutions (4 to 64 m). The model has a reasonably high predictive power as determined by internal validation. The external validation yielded accuracy better than random prediction and we conclude that the existence of residual vegetation is clearly related to the presence of firebreak features and proximity to wet regions. Our repeatable approach is spatially explicit, implemented in an open software environment, and provides acceptable results where local and detailed data availability is substantially limited.