Welcome to Chinese Journal of Applied Ecology! Today is Share:

Chinese Journal of Applied Ecology ›› 2024, Vol. 35 ›› Issue (3): 789-796.doi: 10.13287/j.1001-9332.202403.016

Previous Articles     Next Articles

Prediction of atrazine degradation in soil based on XGBoost model

LI Xiangling1, CHEN Fengxian2, CHEN Xijuan2*   

  1. 1School of Environmental and Safety Engineering, Shenyang University of Chemical Technology, Shen-yang 110142, China;
    2Institute of Applied Ecology, Chinese Academy of Sciences, Shenyang 110016, China
  • Received:2023-08-31 Revised:2024-01-29 Online:2024-03-18 Published:2024-06-18

Abstract: We established the optimal model by using the automatic machine learning method to predict the degradation efficiency of herbicide atrazine in soil, which could be used to assess the residual risk of atrazine in soil. We collected 494 pairs of data from 49 published articles, and selected seven factors as input features, including soil pH, organic matter content, saturated hydraulic conductivity, soil moisture, initial concentration of atrazine, incubation time, and inoculation dose. Using the first-order reaction rate constant of atrazine in soil as the output feature, we established six models to predict the degradation efficiency of atrazine in soil, and conducted comprehensive analysis of model performance through linear regression and related evaluation indicators. The results showed that the XGBoost model had the best performance in predicting the first-order reaction rate constant (k). Based on the prediction model, the feature importance ranking of each factor was in an order of soil moisture > incubation time > pH > organic matter > initial concentration of atrazine > saturated hydraulic conductivity > inoculation dose. We used SHAP to explain the potential relationship between each feature and the degradation ability of atrazine in soil, as well as the relative contribution of each feature. Results of SHAP showed that time had a negative contribution and saturated hydraulic conductivity had a positive contribution. High values of soil moisture, initial concentration of atrazine, pH, inoculation dose and organic matter content were generally distributed on both sides of SHAP=0, indicating their complex contributions to the degradation of atrazine in soil. The XGBoost model method combined with the SHAP method had high accuracy in predicting the performance and interpretability of the k model. By using machine learning method to fully explore the value of historical experimental data and predict the degradation efficiency of atrazine using environmental parameters, it is of great significance to set the threshold for atrazine application, reduce the residual and diffusion risks of atrazine in soil, and ensure the safety of soil environment.

Key words: XGBoost model, prediction, atrazine, soil, degradation