Quantile regression as a generic approach for estimating uncertainty of digital soil maps produced from machine-learning

Abstract

Digital soil mapping (DSM) techniques have provided soil information that has revolutionized soil management across multiple spatial extents and scales. DSM practitioners have been increasingly reliant on machine-learning (ML) techniques; yet, methods to generate uncertainty maps from ML predictions are limited. To address this issue, this study integrates ML-based DSM with quantile regression (QR) in a methodological framework for estimating uncertainty. We test the proposed framework on two case study areas in Canada: (1) a dry-forest ecosystem in the Kamloops region of British Columbia, Canada; and (2) an agricultural system in the Ottawa region of Ontario, Canada. Four ML techniques (Random Forest, Cubist decision tree, k-nearest neighbors, and support vector machine) were compared using repeated cross-validation. Maps showing the 90% prediction interval (PI) were produced. Regardless of the case study, ML approach, and predicted soil variable, the uncertainty estimates were reliable and stable, according to the PI coverage probability analysis.