Assessing the influencing factors of soil susceptibility to wind erosion: A wind tunnel experiment with a machine learning and model-agnostic interpretation approach

Abstract

Soil wind erosion is a worldwide environmental concern. This study explored the feasibility of using a machine learning (ML) algorithm with a model-agnostic interpretation approach to infer the mechanism by which multiple features influence the soil susceptibility to wind erosion. Arenosols were collected from the Ulanbuhe Desert, China, and soil samples with different treatments, including wet-dry and freeze-thaw processes, were used in a wind tunnel experiment and provided a basic dataset for modelling. The Shapley additive explanations (SHAP) method was used to provide global and local interpretations for the ML model. The results showed the following. (1) The mean particle size tended to decrease and the soil organic carbon (SOC) and soil calcium carbonate (CaCO3) contents tended to increase in the transition from mobile to fixed sand dunes. (2) The ML models presented a better performance than the interpretable models, and the random forest model (R2 value of 0.898 on the test dataset) was adopted. (3) All the features except the SOC content showed monotonic or nearly monotonic relationships with the wind erosion rate. Relatively low SOC contents and high particle sizes tended to show positive interaction effects with wet-dry and freeze-thaw processes. An SOC content of 2.46 g kg−1 was deemed a threshold, above which SOC worked to resist wind erosion. ML algorithms and model-agnostic interpretation methods have the potential to improve the reliability and flexibility of inferring the mechanism by which soil is susceptible to wind erosion and thus deserve popularization in future studies.