Exploração de arquiteturas de redes neurais convolucionais para identificação de forrageiras do gênero Bachiaria e Panicum

Abstract

Brazil is one of the world's largest meat exporters due to the low cost of production and mainly to the predominant exploitation in pastures, a fact that makes the country competitive in the international market. It is estimated that in Brazil the total coverage area with cultivated pastures is 100 million hectares. Pastures are considered the cheapest and main source of food for cattle raising. The cultivars of two tropical genera are highlighted in the Brazilian seed market: Brachiaria and Panicum. Brachiaria is the most used, adapts to various soil and climate conditions and has great tolerance to weak and acidic soils. It shares space with Panicum, which, unlike Brachiaria, is recommended for more fertile soils. These two genera are the basis for studies of several Embrapa programs, which aim to launch more and more new cultivars. Other programs involve mapping pasture areas. Identify cultivars planted in different regions of Brazil. However, there are difficulties, even by specialized technicians, to identify the name, species and genus of the plant. During the dry and rainy seasons, the plants undergo morphological changes, which can make it even more difficult. The hierarchical classification of each forage follows standards of biotaxonomy, a technique responsible for naming plants. This hierarchy should classify the plant, first by name of the cultivar, then by species and, finally, by genus. In this context, this work aims to explore the architectural capacity of convolutional neural networks to identify sixteen forages per image, at the level of classification by Cultivar, Species and Gender. Considering the physical changes of plants, during the dry and rainy seasons. Another important issue was to contribute to forming an image bank of these two types of forage. The images were collected at Embrapa Gado de Corte, in Campo Grande - MS. Therefore, the images that were taken from june to november 2019 made up the drought period dataset, while the images that were taken between december 2019 and february 2020 made up the rainy season dataset. Convolutional neural networks are applied with great success in image recognition. Proof of this is the constant emergence of new state-of-the-art architectures. The project explores four convolutional network architectures, two state-of-the-art, MobileNet and ResNet50 and two others assembled according to the literature, called CNN I and CNN II. Cultivar classification accuracy was the lowest. As for species and genus, they were the best, demonstrating that convolutional networks have the potential to distinguish forages by species and genus. State-of-the-art architectures achieved the best results. The differences in the performance of the nets, in both periods, were small; not allowing to affirm that, classifying forages in the rainy season is easier than in the dry season and vice versa.