Provincial-scale digital soil mapping using a random forest approach for British Columbia

Abstract

Although British Columbia (BC), Canada, has a rich history of producing conventional soil maps (CSMs) between 1925 and 2000, the province still lacks a detailed soil map with a comprehensive coverage due to the cost and time required to develop such a product. This study builds on previous digital soil mapping (DSM) research in BC and develops provincial-scale maps. Soil taxonomic classes (e.g., great groups and order) and parent material classes were mapped at a 100 m spatial resolution for BC (944 735 km2). Training points were generated from detailed and semi-detailed soil survey maps. The training points were intersected with 26 topographic indices for mapping parent materials with an additional 9 climatic and vegetation indices for mapping soil classes. The soil–environmental relationships were inferred using the random forest (RF) classifier. The fitted models were used to predict 23 soil great groups, 9 soil orders, and 10 parent material classes. Accuracy assessments were performed using n = 14 570 validation points for parent material classes and n = 14 316 validation points for soil classes, acquired from the BC Soil Information System. The accuracy rates for soil great groups, orders, and parent material classes were 55%, 62%, and 69%, respectively, and kappa coefficients were 0.37, 0.41, and 0.59, respectively. This study demonstrated that when RF was trained using CSMs, the accuracy for the resulting DSM was higher than the original CSM. To assess prediction uncertainties, ignorance uncertainty maps were developed using class-probability layers generated by the RF models.