Classification of soybean cultivars by means of artificial neural networks

Resumo

The cultivation of soy has an economic importance for the Brazilian agricultural scenario. The aim of this study was to establish a network architecture for the classification of soybean genotypes, by means of morphological characters measured in the juvenile phase of the plant, and finally to compare the results obtained through Artificial Neural Network (ANN) and Anderson Discriminant Analysis. The study analyzed plants of 10 conventional cultivars in the initial stages of development (V1, V2 and V3 stages). The experiment was carried out in a randomized block design with 5 replications, and the experimental unit was represented by 9 plants. The data were submitted to the Anderson Discriminant Analysis and multilayer Perceptron ANN, with 1 or 2 hidden layers. To analyze the homogeneity of the variance and covariance matrix, the Box’s M-Test was adopted in the Program R, at 5% significance level. An input layer, one or two hidden layers, and an output layer formed the ANN architecture. The 5-fold cross validation was used to verify the efficiency of the discriminant functions and also in the ANN analysis. Subsequently, the apparent error rate (AER) was obtained. Box’s M-Test indicated inhomogeneity in the variance and covariance matrices, which indicated the need to perform Anderson's Quadratic Discriminant Analysis. The ANNs presented lower apparent error rate when compared to the Anderson's Quadratic Discriminant Analysis and the artificial neural network with 1 hidden layer was sufficient to perform the classification of soybean cultivars.