Predictive modeling of diazinon residual concentration in soils contaminated with potentially toxic elements: a comparative study of machine learning approaches

The widespread use of pesticides, including diazinon, poses an increased risk of environmental pollution and detrimental effects on biodiversity, food security, and water resources. In this study, we investigated the impact of Potentially Toxic Elements (PTE) including Zn, Cd, V, and Mn on the degradation of diazinon in three different soils. We investigated the capability and performance of four machine learning models to predict residual pesticide concentration, including adaptive neuro-fuzzy inference system (ANFIS), support vector regression (SVR), radial basis function (RBF), and multi-layer perceptron (MLP). We employed a 10-fold cross-validation mechanism to evaluate the models. Moreover, performance validation of selected algorithms through the coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE) and mean square error (MSE) confirm that the SVR and ANFIS with lower RMSE, MSE, and a higher R² can simulate the degradation process better than other models. The result showed that both SVR and ANFIS approaches worked well for the data set, but the SVR technique is more accurate than the fuzzy model for estimating pesticide concentration in soil in the presence of PTE. Vanadium appeared to be the best option for the degradation of diazinon. The models predicted the performance of V²⁺ for diazinon degradation with R² and RMSE of 0.99 and 2.18 \(mg.kg^{-1}\) $mg.k{g}^{-1}$ for SVR and, 0.99, and 1.30 for the ANFIS model for the training set. Finally, the high accuracy of the models was confirmed.