Artificial intelligence for herbicide recommendation: Case study for the use of clomazone in Brazilian soils

• The use of RNAs was tested to estimate the sorption capacity of clomazone; • The RNA-MLP models were efficient in determining the sorption coefficient using the physical and chemical properties of the soil; • The feature selection method proved to be a great tool for predicting networks; • ANNs allowed recommendations of low doses for soils with reduced sorption and high doses for soils with greater clomazone sorption; • The K + and Silt content were the main soil attributes with the greatest importance in predicting the sorption coefficient. The use of herbicides has been carried out without considering the physical and chemical properties of the soil and its variation in the different producing regions. Its interaction is dependent on the properties of the soil and the herbicide itself and, therefore, is considered complex. An alternative to optimize the use of herbicides is the application of mathematical models to estimate the sorption in the soil, contributing to a more efficient control of weeds and bringing more environmental safety. In this study, the efficiency of Artificial Neural Networks (ANNs) was evaluated as a tool to predict clomazone sorption in different Brazilian soils. The herbicide sorption coefficients were determined in the laboratory with 45 soils from different soil environments. Multilayer perceptron (MLP) ANN models were used to predict clomazone sorption. The variables were selected using the feature selection tool using the physical and chemical properties of the soils. Multivariate statistics were performed to determine the correlations between the physical and chemical properties of the soil with the sorption coefficient. Potassium (K), phosphorus (P), magnesium (Mg), organic matter (OM), silt, clay, and cation exchange capacity (CEC) were the ANN inputs and the sorption coefficient (Kfs) the output. The general performance of the model was evaluated by its precision and error values, namely: coefficient of determination ( R 2 ), mean absolute relative error ( RMSE ), mean absolute error ( MAE ), mean estimation error ( MBE ), and the coefficient Pearson correlation test ( r ). The ANN models were able to predict the Kfs of clomazone in the studied soils. The input variables that determined the best performing network for Kfs prediction were: CTC, Silt, Mg, and K. The variables that showed greater relative importance in the sensitivity analysis in the construction of the model for the Kfs of clomazone were: K (37%) and Silt (29%). The best RNA model was able to recommend reducing the dose of clomazone compared to the commercial package insert method.