Predição da necessidade de calagem em área de pivô central por meio de rede neural artificial

Abstract

Predicting The Need For Liming In A Central Pivot Area Through An Artificial Neural Network. Much of the state of Goiás is composed of latosols, due to the high occurrence of acidity in this type of soil, it is essential to study corrective methods and new technologies to assist in the management of these soils, as well as artificial neural networks (ANN). which has been shown to be viable for studies of soil chemical attribute estimates. The objective of this work is to estimate the need for liming in central pivot areas using artificial neural networks. The data were collected in an area of 35 ha, in a 60x60 m sampling grid, generating 88 sample points. For each of these points, 5 subsamples will be collected in the 0 to 0.20 m layer to determine the chemical attributes and soil texture. The data were submitted to descriptive and exploratory analysis and soon after they were normalized for insertion in the ANNs, so that they could be calibrated, then their estimates were compared to the sample data the need for liming, to verify the network performance, by means of the agreement index, correlation coefficient and verification of the lowest mean relative error. In the two models involved, the ANNs were able to make the proposed estimates with excellent performance. Various structures were tested and according to the statistical indicators, the best network for each model was arrived at. In model 1, RNA1 (Artificial Neural Network with 2 neurons in the hidden layer) achieved the best results, presenting an excellent performance (id = 0.993) and the mean quadratic error (MSE) of 0.002. In model 2, RNA 1 (Artificial Neural Network with 6 neurons in the hidden layer) achieved the best results, presenting an excellent performance (id = 0.929) and the mean quadratic error (MSE) of 0.025. The t-test proved the accuracy of the two models, since there were no significant differences in the estimates using ANNs and the calculation by the traditional method of liming requirement. Model 2, using the neighborhood technique adapted for data insertion in ANN, managed to decrease the number of samples to be collected by 30% for a possible mapping of the need for liming at a varied rate.