Too Sour to be Lime: Improving Consistency of Digital Soil Mapping with multivariate neural network and Soil Science Informed Loss Constraint.

Abstract

Machine learning is increasingly used in the generation of maps from discrete point measurements, particularly often in Digital Soil Mapping. Commonly, the interdepen dence between soil properties and constraints of sparse and limited soil data is not ad dressed, leading to unrealistic predicted combinations of soil properties. Potential so lutions include the use of multivariate models and incorporation of domain knowledge through techniques proposed by physics-informed ML. This thesis examines whether using a multivariate neural network and adding soil science informed loss term improves prediction consistency without sacrificing accuracy, focusing on soil pH and calcareous content across arable land in Zurich, Switzerland. The findings suggest that the use of multivariate models can increase the consistency without reducing accuracy. The in corporation of a penalty term for the relationship between outputs is possible and has the potential to effectively enforce the known relationship without sacrificing accuracy. However, the practical implementation of the penalty requires careful tuning and its utility is likely larger in more data-scarce scenarios. In this application the multivariate model alone was sufficient to infer the underlying relationships between soil properties, presumably due to imputation-enriched data.