Modeling Rapidly Discriminative Strategies of Cr Contaminated Soils Through Machine Learning

Abstract

Chromium contamination in soil is a major threat to water and ecosystems. However, there has been a lack of methods to rapidly identify strategies for treating Cr-contaminated soils. In this study, We developed 12 machine learning (ML) models to predict the washing treatment result using 250 experimental data points with consideration of the main control factors (soil properties, Cr sequential extraction and type of washing agent). The results of model performance comparison showed that the lightGBM model performed best for predicting the Cr content in the soil after washing treatment to confirm whether the Cr contaminated soil is suitable for a washing release strategy, and the Adaboost model performed best for most of washing agents. Based on the feature analysis, soil pH, exchangeable potassium, reactive Fe oxides, and Cr sequential extraction can account for most of the differences in washing treatment. The results of the model application showed that the error between the predicted and actual test values of the soil leaching treatment results for two brand new soils was less than 5%. This study will rapid and accurate prediction of the strategy of Cr contaminated soil treatment and significantly improve the efficiency of technological remediation of Cr contaminated soils.

Keywords: Cr-contaminated soils, Washing, Machine learning (ML), Feature analysis, Remediation