Classificação da viabilidade de sementes de soja com uso de inteligência computacional

Abstract

The classification of soybean seed lots makes it possible to obtain information about which seed lot has the greatest potential to germinate and quickly generate normal seedlings under adverse environmental conditions. The use of tools such as computational intelligence contributes to the differentiation of soybean seed lots quickly, precisely and accurately. Thus, the hypothesis of this work is based on the possibility of classifying soybean seeds from four years of analysis in relation to the physiological quality presented. Therefore, the objective of this work is to evaluate the viability classification of soybean seeds through computational intelligence. The data used for machine learning were obtained at the Seed Production and Technology Laboratory at the Federal University of Mato Grosso Do Sul – CPCS/UFMS, through a survey of the results of four years of soybean seed analysis. The results were on cards, containing information about the first germination count (PCG), germination, tetrazolium vigor (vigor) and tetrazolium viability (viability). After obtaining the data referring to the analyzes carried out in seeds, these were tabulated and submitted to statistical analysis. The data were subjected to principal component analysis (PCA) associated with the k-means algorithm, forming four clusters. Subsequently, the data were submitted to machine learning analysis, where the formed clusters were used as output variables and the analyzes carried out to evaluate the physiological quality of soybean seeds were used as input (input). The performance of the classification models used was evaluated by the percentage of correct classifications (CC), Kappa and F-score metrics. According to the grouping, boxplots were used with the means of the accuracy models, with means compared by the Scott-Knott test at 5% significance. The classification of soybean seed quality over four years of evaluation was performed efficiently by the RL, RNA and SVM algorithms. Such algorithms obtained accuracy metrics above 70%.