Improved Soil Mapping in British Columbia, Canada, with Legacy Soil Data and Random Forest

Abstract

The need for improved soil inventory information in the province of British Columbia (BC), Canada, was addressed using a random forest (RF) classifier that was informed using legacy soil data. RF models were prepared for 110 ecodistrict subdivisions of BC, and predictions were subsequently assembled into a final soil parent material map mosaic covering the entire province. The ecodistricts are part of a framework for ecosystem classification in BC and in Canada, and delineate areas with relatively homogeneous biophysical and climatic conditions. Training areas for predicting soil parent materials were identified using single-component polygons from legacy terrain, soil, and ecosystem maps. For parent material mapping, we intersected training points amalgamated from all legacy surveys with a suite of 18 topographic covariates derived from a 100-m digital elevation model (DEM). For each ecodistrict, two versions of the resulting training dataset were submitted to the RF classifier. A ‘balanced’ dataset contained equal numbers of training data points for all parent material classes representing all legacy data derived from single-component polygons. A ‘constrained’ dataset was also derived where conditions were imposed on selected topographic attributes of the training points to reflect known geomorphic processes and to ensure consistent mapping criteria were applied across multiple legacy soil survey projects. RF predictions of soil parent material resulted in 100-m gridded class maps for BC that incorporate expert knowledge extracted from legacy soil inventories.