Soil contamination by persistent agrochemicals such as tebuthiuron (TB) and thiamethoxam (TX), together with the use of agroindustrial by-products like vinasse (VN), affects soil-plant interactions and crop development. This study evaluated the biometric responses of Pennisetum glaucum grown by measuring root and shoot dry biomass and leaf chlorophyll index in soils treated with combinations of TB, TX, and VN under controlled greenhouse conditions. Nine regression models were tested to predict plant responses based on these variables, considering or not plant development time, and interpreted using Shapley Additive Explanations (SHAP). The Gaussian Process Regression (GPR) and Random Forest (RF) models achieved the highest predictive performance (R² up to 0.27), whereas model accuracy declined sharply (R² < 0.05) when time was excluded, confirming its critical influence on predictive reliability. Vinasse improved plant growth, while TB and TX exerted neutral or negative effects, reflecting their persistence and potential phytotoxicity. These findings demonstrate the capacity of interpretable machine learning to elucidate soil-plant-contaminant interactions and support the design of sustainable management strategies for pesticide-impacted agricultural soils.
Frias et al. (Mon,) studied this question.