Abstract: Machine learning has been widely used in the prediction of soil physical, chemical and biological properties in recent years. We summarized the research advances in this area from the following three aspects: (1) Typical machine learning types and common machine learning algorithms in prediction of soil property, including random forest, support vector machine, and BP neural network; (2) Model accuracy indices commonly used in regression prediction of soil property; (3) The general process and recent research cases of soil property prediction using machine learning. In the future research of soil property prediction, we should pay attention to the uncertainty assessment of prediction and coupling multiple models through ensemble methods, which would help improve work efficiency and pursue high accuracy of models. It is necessary to explore ways to improve the interpretability, generalization and robustness of models. When selecting machine learning algorithms and comparing models, it is recommended to use multiple types of predictor variables to ensure the accuracy of prediction after consideration of research needs and content.

Key words: soil property, machine learning, ensemble learning, deep learning, digital soil mapping, soil health