Enhancement of the use of machine learning in digital soil mapping

PhD project Stephan van der Westhuizen

The use of machine learning in digital soil mapping (DSM) is becoming increasingly popular, because it can easily encapsulate linear- and nonlinear relationships and produce accurate predictions. The flexibility of machine learning models provides opportunities to enhance its use in DSM even further. This project will address four important problems in machine learning for DSM: (i), Measurement error in the soil property, that is, how to incorporate measurement error in calibration data into a machine learning model to improve model calibration and prediction; (ii) Incorporation of expert knowledge into a model. How to incorporate expert knowledge about soil properties with machine learning models; (iii) How to perform multivariate soil property prediction with machine learning models; (iv) How to model compositional soil data with machine learning models. Methodological innovations will be tested using synthetic and real-world soil data from the WOSIS and SoilGrids soil information system.