Utilização de inteligência artificial para a detecção de plantações em imagens de satélite

• Ano de defesa: 2022
• Autor(a) principal: Jacques, Matheus Mello
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Santa Maria; Brasil; Engenharia Elétrica; UFSM; Programa de Pós-Graduação em Engenharia Elétrica; Centro de Tecnologia
• Palavras-chave em Português: Detecção de plantações; Redes neurais; Sentinel-2; U-Net; Visão computacional; Crop type detection; Neural networks; Computer vision; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA
• Link de acesso: http://repositorio.ufsm.br/handle/1/28173

Loss identification is one of the main challenges for electric utility companies today. The loss factor in Brazil was approximately 14% in 2018, where 6.6% are non-technical losses totaling a cost of almost R$5 billion per year. This work proposes the location of thefts in the electrical network using satellite images. For this, Deep Learning models are used to identify rice crops and later this data feeds another artificial intelligence algorithm that analyzes whether such consumer unit is generating non-technical losses. This work, therefore, proposes the use of the topology proposed by (??) for the classification of plantations in the Uruguaiana region in Rio Grande do Sul. Sentinel-2 satellite images were acquired by the (ESA, 2021) platform, processed and used as input to the neural network. In addition, they were also used for the creation of the labels, through the use of the Maximum Likelihood method combined with the visual inspection of a professional. After training the neural network for 30 epochs, we found a loss function close to 0.3 in the test dataset, an accuracy of 90% and a Jaccard Index of 68.84%. Through the application of this Deep Learning model, it was possible to detect and classify the plantations, to then cross-reference with data from the electrical network to identify possible cases of losses.