Use of neural networks to estimate the sorption and desorption coefficients of herbicides: A case study of diuron, hexazinone, and sulfometuron-methyl in Brazil.

The use of herbicides in Brazil has been carried out based on the manufacturer's recommendation, often disregarding the high variability of soil attributes. The use of statistical methods to predict the herbicide retention processes in the soil can contribute to the improvement of weed control efficiency associated with the lower risk of environmental contamination. This research evaluated the use of Artificial Neural Networks (ANNs) to predict soil sorption and desorption, as well as the environmental contamination potential of diuron, hexazinone and sulfometuron-methyl herbicides in Brazilian soils. The sorption and desorption coefficients of the three herbicides were determined in laboratory tests for 15 soils from different Brazilian states. To predict the sorption and desorption of diuron, hexazinone and sulfometuron-methyl were used a multilayer perceptron ANNs (MLP). The inputs were the characteristics of the herbicides and the physical and chemical attributes of the soils, and the outputs of were the sorption and desorption coefficients (Kfs and Kfd). The risk of leaching of diuron, hexazinone, and sulfometuron-methyl herbicides were evaluated considering the sorption values observed and those estimated by the models. The Artificial Neural Network (ANN) models were efficient for the prediction of sorption and desorption of diuron, hexazinone, and sulfometuron-methyl herbicides. The physicochemical properties of the herbicides were more important for the modeling of multilayer perceptron ANNs than the soil attributes. The herbicides diuron, hexazinone, and sulfometuron-methyl have a high potential risk for contamination of groundwater in different Brazilian states.