Estimativa da necessidade de fósforo e potássio para o tomate industrial utilizando redes neurais artificiais

Abstract

In agriculture, it is possible to find the application of artificial neural networks (ANNs) in studies aimed at predicting attributes of crop productivity, among others, and given the breadth of use, ANNs consist of a promising method to estimate indicators related to soil quality. This study aimed to develop an ANN to estimate the needs of phosphorus (P) and potassium (K) in central pivot areas cultivated with industrial tomatoes and to reduce the amount of samples required for field data collection. The data were collected using a sample grid of 60x60m, totaling 88 sampling points. Two network models were trained, the first aiming at estimating the need for phosphorus and the need for potassium in central pivot areas and the second aiming at reducing the amount of samples needed for data collection in the field, in which the back-Propagation algorithm was used for the training, with Multiple Layer Perceptron (MLP) topology. Subsequently, a geostatistical analysis was performed and the semivariogram model was adjusted for the P and K requirement data estimated by the fertilizer recommendation table, based on the soil analysis, and those estimated by the ANN model 2, in which the maps were prepared. of isolines and evaluated the accuracy of the maps through the Kappa and Global Accuracy indices. For the two trained models, good adjustments were observed with R2 values greater than 0.90 in the external validation phase, low EQM values (mean squared error), correlation above 91% and excellent performance of the ANNs, in which the precision of the chosen networks were confirmed by the t-Test, for both trained models. It was possible to estimate the needs of P and K and the adaptation of the neighborhood technique enabled a 40% reduction in the number of samples needed for the collection of field data, making it possible to obtain estimates of the needs of P and K, aiming at a possible application of nutrients at a variable rate, contributing to cost reduction. The Kappa index and global accuracy allowed the evaluation of the accuracy of the prepared maps, in which medium and high accuracy was found between the P and K needs maps, respectively, and global accuracy values greater than 85%. By reducing the number of samples, the network is able to obtain valid estimates of P and K to be used in the study of soil spatial variability.