Explainable machine learning for arsenic remobilization potential in the vadose zone: Leveraging readily available soil properties

Abstract

The vadose zone acts as a natural buffer that prevents contaminants such as arsenic (As) from contaminating groundwater resources. Despite its capability to retain As, our previous studies revealed that a substantial amount of As could be remobilized from soil under repeated wet-dry conditions. Overlooking this might underestimate the potential risk of groundwater contamination. This study quantified the remobilization of As in the vadose zone and developed a prediction model based on soil properties. 22 unsaturated soil columns were used to simulate vadose zones with varying soil properties. Repeated wet-dry cycles were conducted upon the As-retaining soil columns. Consequently, 13.9-150.6 mg/kg of As was remobilized from the columns, which corresponds to 37.0-74.6 % of initially retained As. From the experimental results, a machine learning model using a random forest algorithm was established to predict the potential for As remobilization based on readily accessible soil properties, including organic matter (OM) content, iron (Fe) content, uniformity coefficient, D30, and bulk density. Shapley additive explanation analyses revealed the interrelated effects of multiple soil prop-erties. D30, which is inter-related with Fe content, exhibited the highest contribution to As remobilization, fol-lowed by OM content, which was partially mediated by bulk density.