A new alternative to determine weed control in agricultural systems based on artificial neural networks (ANNs)

Weed control is a necessary practice to avoid crop yield losses. Therefore, farmers should answer the following question: when to start weed control? Currently, there are no learning models to assist the producer to answer this question. Thus, the objectives were to: 1) evaluate the ability of artificial neural networks (ANNs) to estimate the beginning of weed control for different classes of acceptable yield losses; 2) validate a new alternative for modeling and predicting competition between weeds and crops. ANNs determined the ideal moment to control weeds based on non-destructive and destructive variables. The inputs C3/C4 ratio, coexistence period, density of weeds, and crop (categorical variable to differentiate sesame and melon) provided accuracy and F-score values above 0.95 during training, validation, and testing steps for ANN in non-destructive method. When using the destructive variables, C3/C4 ratio plus coexistence period, fresh matter of weeds, and crop provided accuracy and F-score values above 0.90 during training, validation, and testing steps. The combination of non-destructive and destructive inputs also generated an ANN with high accuracy and F-score, above 0.95, during training, validation, and testing steps. Machine learning can be used in crop-weed competition modeling.