Semi-automated classification of exposed bedrock cover in British Columbia's Southern Mountains using a Random Forest approach

Knowledge of the spatial distribution of exposed bedrock (EB) is essential for natural resource inventories, environmental monitoring, and landscape evolution modelling. This paper presents a method for the use of a Random Forest (RF) classifier and legacy land data to locate areas of EB in a mountainous landscape of southern British Columbia, Canada. EB map accuracy increased from 48% to 88% with the use of RF models in comparison to the legacy land cover maps. Reducing the total number of predictor variables from 43 to 17 had a negligible effect on prediction accuracy. Meaningful relationships between EB predictions and important predictor variables were observed in partial dependence plots. These findings emphasize that substantial improvement in EB map accuracy is possible with a limited number of predictor variables, and interpretation of such maps is improved where the relationships between specific predictor variables and predicted classes can be observed and evaluated.