Incorporating soil knowledge into machine-learning prediction of soil properties from soil spectra

Abstract

Various machine-learning models have been extensively applied to predict soil properties using infrared spectroscopy. Beyond the interpretability and transparency of these models, there is an ongoing discussion on the reliability of the prediction of soil properties generated from soil spectra. In this review, we contribute to this discussion by advocating for the integration of soil knowledge into machine-learning models. By doing so, researchers can delve deeper into the underlying soil constituents, ultimately enhancing prediction accuracy. Our review explores the soil information present in spectral data, the fallacy of model interpretability, methods to incorporate soil knowledge into machine-learning techniques, and the ways in which machine learning and soil spectroscopy can assist soil science. The combination of machine learning and domain knowledge is recommended to develop more meaningful models for predicting soil properties within the field of soil science.