Detecção de linhas de plantio em plantações de cana-de-açúcar utilizando deep learning

• Ano de defesa: 2023
• Autor(a) principal: Ribeiro, João Batista
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Ciência da Computação
• Palavras-chave em Português: Linhas de Plantio; Deep Learning; Agricultura de Precisão; VANTs; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://repositorio.ufu.br/handle/123456789/39346; http://doi.org/10.14393/ufu.di.2023.8085

Rapid population growth has driven demand for food and the sustainable use of natural resources. In this context, agriculture allied to technology, called Precision Agriculture (PA), seeks to meet this demand using tailored resources based on the information collected. The images used in the PA have different sources, e.g., cameras attached to Unmanned Aerial Vehicles (UAVs). One of the main applications of PA is the detection of planting lines, mainly because it is an important step for other PA applications, e.g., weed detection, crop production mapping and forecasting, fault detection. One of the scenarios of great use of PA in Brazil is in the cultivation of sugarcane, motivating researchers and companies to develop solutions in the area. Although there are many works in the literature to detect planting lines, most of them are for other crops, e.g., maize and beet, or focused on straight lines. Considering the scenario of great use of UAVs to obtain images for PA and the great importance of detecting planting lines, this project analyzed different Deep Learning (DL) models for automatic segmentation in UAV images of sugarcane plantations with varying stages of growth. Among the models, U-Net achieved the best results, with 0.90 or more Dice Coefficient (DC) for almost all scenarios. The use of Vegetation Indexes (VIs) and Morphological Operations was also analyzed in order to optimize the detection of planting lines. Based on the results, some recommendations are presented for using U-Net and VIs to obtain greater precision in the segmentation of UAVs images of sugarcane plantations.