ARTIFICIAL NEURAL NETWORKS TO ESTIMATE THE SORPTION AND DESORPTION OF THE HERBICIDE LINURON IN BRAZILIAN SOILS.

Generally, herbicides used in Brazil follow manufacturer's recommendations, which often do not consider soil attributes. Statistical models that include the physicochemical properties of the soil involved in herbicide retention processes could enable greater precision in herbicide dose decision-making. This study evaluated the potential of artificial neural networks (ANNs) to predict the sorption and desorption of the herbicide linuron in Brazilian soils with different attributes. ANNs and multilayer perceptron (MLP) models were built to predict the sorption and desorption of the herbicide linuron. The inputs to the networks were pH, organic matter (OM), clay, cation exchange capacity (CEC), and base saturation (V); the outputs were sorption (Kfs and Qmax) and desorption (Kfd). The performance of the predictive model was assessed by the coefficient of determination (R), mean absolute relative error (RMSE), mean absolute error (MAE), mean estimation error (MBE), and Pearson's correlation coefficient (r). The best-performing ANNs for predicting Kfs and Kfd comprised the variables pH, OM, CEC, and clay; for predicting Qmax, the ANN comprised the variables pH, OM, and clay. Artificial neural network models have proved to be a valuable tool for predicting the sorption and desorption of the herbicide linuron in soil, helping to minimize environmental impacts by providing accurate estimates and promoting sustainable herbicide use based on soil attributes.