Classificação do estádio de maturação de cagaitas utilizando estatística multivariada e redes neurais artificiais

Abstract

The present work aimed to evaluate the classification of different maturation stages of cagaita (Dysenteric eugenics DC.) using multivariate statistics and artificial neural networks. 120 samples were collected, with 3 fruits each sample, every 3 days on 5 different days after anthesis (DAA) (28, 31, 34, 37 and 40 DAA), totaling 600 samples. For the analysis using multivariate statistics, characteristics were evaluated, such as: weight, longitudinal diameter, transversal diameter, area, color (luminosity, °hue and chroma), pulp and skin firmness, soluble solids (SS), hydrogenion potential (pH), titratable acidity (AT) and maturation index (IM). The data obtained were standardized and submitted to principal components analysis (PCA) and cluster analysis (Clusters), and to validate the established methods, the cophenetic correlation coefficient (CCC) was calculated. In the classification of the cagaita maturation stage using artificial neural networks, images of each cagaita fruit were obtained on each day after the anthesis analyzed, totaling 600 images. The images were converted to grayscale and segmented. Subsequently, information was extracted from the color models (RGB, HSV, L, °hue and chroma) and the geometric attributes (length and area) of the fruit images. The parameters extracted from the images were used as inputs to the artificial neural networks. The outputs were established based on the days after anthesis and analyzed, in which each one came to correspond to a class (1, 2, 3, 4 and 5), respectively. 20 networks were trained, changing, in each of them, the number of nodes in the hidden layer (3 to 12 nodes) and the optimization functions (adam and nadam). The precision and loss (error) of each network were calculated in order to determine which model presented the best performance in the training. To evaluate the performance of the best model, 4 statistical parameters were calculated, namely: accuracy, precision, recall and F-score. For the first stage of the analysis, through principal component analysis, PCs 1, 2 and 3 explained 98,44% of the data variability. Cluster analysis (Clusters) explained the maturation stage of cagaitas through the behavior of physical, physicochemical and chemical variables at different stages of maturation, dividing these days into 3 groups (group 1, group 2 and group 3), where in group 1 we find if 28 and 31 DAA, group 2, 40 DAA and group 3, 34 and 37 DAA. Multivariate statistical analyzes showed a cophenetic correlation coefficient of 0,8234. For the second part, using the ANNs, the network model that presented the best precision was the one with 9 nodes in the hidden layer and “adam” optimizer, showing the accuracy of 57,353%. In the network classification metrics report for determining different stages of maturation, the highest accuracy, precision and R-score are from Class 1, 97,1%, 75% and 75%, respectively, and the best recall from Class 2 (88,9%). Checking the calculation of the cophenetic correlation coefficient, regarding the adequacy of the grouping and the precision of definition of the best model of ANNs, it was possible to conclude that the best method for the classification of different stages of maturation of cagaita in this work was the multivariate statistics, although the classification through artificial neural networks was also efficient.