Detecção de painéis fotovoltaicos em ortofotos utilizando redes neurais profundas

• Ano de defesa: 2021
• Autor(a) principal: Antônio Campos de Abreu Filho
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; IGC - DEPARTAMENTO DE CARTOGRAFIA; Programa de Pós-Graduação em Análise e Modelagem de Sistemas Ambientais; UFMG
• Palavras-chave em Português: energia solar fotovoltaica; aprendizado profundo; segmentação de instâncias; geoprocessamento; Modelagem de dados – Aspectos ambientais; Energia solar; Geração de energia fotovoltaica – Belo Horizonte (MG); Geoprocessamento
• Link de acesso: http://hdl.handle.net/1843/40975

Mapping spatial information has increasingly gained relevance in urban planning and city management applications. The automated analysis and classification of images have been a problem in several areas of knowledge and the remote sensing community has shown particular interest in the application of techniques that use deep neural networks in the land use segmentation task. To train these networks, extensive labeling is required for the objects to be identified. Instance segmentation, in particular, requires even more complex and laborious labeling. The world demands energy and "Swanson’s Law" predicts that photovoltaic modules will achieve grid parity in a few years. Compared to other alternatives, rooftop photovoltaic generation is more widely accepted for using idle spaces in urban buildings. Studying how photovoltaic installations evolve in urban areas is an important indicator of the evolution of energy security in cities and its detection based on aerial images is an important way of understanding it. This work aimed to train a deep neural network using publicly available data from Rio de Janeiro/RJ, Brasília/DF and Campinas/SP. Using the trained model, an orthophoto, covering the entire area of Belo Horizonte/MG, was segmented, on a 7 cm/pixel spatial resolution, evaluating all photovoltaic modules and solar heating systems in the study area. 1719 manually labeled data were used as training samples and the inference step identified 3655 photovoltaic modules and 26369 solar heating systems in the area. As validation metrics of the model, the following values were obtained: global average precision (AP=0.18), global average recall (AR=0,36), average precision of solar panel’s class (APUFV=0,30) and average precision of solar heating systems class (APSAS=0,16).