Digital mapping of soil properties in Canadian managed forests at 250m of resolution using the k-nearest neighbor method

Publication date: December 2014
Source:Geoderma, Volumes 235–236
Author(s): Nicolas Mansuy , Evelyne Thiffault , David Paré , Pierre Bernier , Luc Guindon , Philippe Villemaire , Vincent Poirier , André Beaudoin
Large-scale mapping of soil properties is increasingly important for environmental resource management. While forested areas play critical environmental roles at local and global scales, forest soil maps are typically at low resolution. The objective of this study was to generate continuous national maps of selected soil variables (C, N and soil texture) for the Canadian managed forest landbase at 250m resolution. We produced these maps using the kNN method with a training dataset of 538 ground-plots from the National Forest Inventory (NFI) across Canada, and 18 environmental predictor variables. The best predictor variables were selected (7 topographic and 5 climatic variables) using the Least Absolute Shrinkage and Selection Operator method. On average, for all soil variables, topographic predictors explained 37% of the total variance versus 64% for the climatic predictors. The relative root mean square error (RMSE%) calculated with the leave-one-out cross-validation method gave values ranging between 22% and 99%, depending on the soil variables tested. RMSE values<40% can be considered a good imputation in light of the low density of points used in this study. The study demonstrates strong capabilities for mapping forest soil properties at 250m resolution, compared with the current Soil Landscape of Canada System, which is largely oriented towards the agricultural landbase. The methodology used here can potentially contribute to the national and international need for spatially explicit soil information in resource management science.