Modelos computacionais baseados em aprendizado de máquina para classificação e agrupamento de variedades de tucumã (Astrocaryum aculeatum G. Mey.)

Abstract

The amazon biome has a range of natural resources with high economic value, which can be exploited in a sustainable way to generate jobs and income. Among these resources, we can spotlight the tucuman, native fruit from the Amazonian region used in cooking, crafts and sold by local producers. The tucuman palm tree presents many varieties within the same specie, each one differs from the other in its morphology, population, origin, among other factors. Scientifically, the tucuman taxonomic classification refers to its species. There isn’t yet a manual or an automatic way of classifying varieties of Astrocaryum aculeatum G. Mey., also known as Amazonian tucuman. The indication of the variety to which a fruit unit belongs is performed empirically and may be confused when there are fruits with a large difference in their characteristics. In this scenario, this study aimed to generate and evaluate computer models able to classify and get into groups four varieties of tucuman found in the Itacoatiara-AM region. The secondary objective of this study was to indicate which of the varieties have the best economic potential regarding to the harvested fruit characteristics. To generate the models, three machine learning techniques were used: Decision Trees and Artificial Neural Networks in the classification task, and to the grouping task the K-Means technique was applied, using Euclidean and Manhattan distance measurements. The results obtained based on the data set show that the models generated with machine learning techniques presented satisfactory indexes for predicting of varieties' classes of tucuman.