Detecção de fogo ativo por aprendizado profundo em imagens provenientes do satélite Landsat-8

• Ano de defesa: 2021
• Autor(a) principal: Fusioka, Andre Minoro
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Curitiba; Brasil; Programa de Pós-Graduação em Computação Aplicada; UTFPR
• Palavras-chave em Português: Fogo - Detecção; LANDSAT (Satélites); Algorítmos computacionais; Redes neurais (Computação); Fire - Detection; LANDSAT satellites; Computer algorithms; Neural networks (Computer science); CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO; Ciência da Computação
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/26647

Active fire detection in satellite imagery is of great importance for the development of environmental conservation policies, supporting decision-making and law enforcement. Active fire detection techniques are generally based on comparing pixels or image regions with specific thresholds for the used sensor. This work addresses active fire detection using deep learning techniques. In recent years, deep learning techniques have enjoyed enormous success in various fields, but their use for active fire detection is relatively new, só it is necessary to analyze the feasibility of using deep learning for active fire detection. This study evaluates how different convolutional neural network architectures can be used to segment active fire in satellite images, using the masks produced by methods commonly used to detect active fire as training samples, as well as the combination of such masks, evaluating images from the Landsat-8 satellite from the period of August 2020. In addition, the studied models are evaluated on images from the same satellite obtained in September 2020, but manually annotated, verifying the capacity of the networks to approximate annotations made by a human specialist. The tested architectures were able to approximate the classical methods of the literature, and the best overall observed performance obtained an -score metric of 94.2% and an IoU of 89.0%. When compared with the manual annotations, the networks obtained results superior to the classic methods of the literature in the vast majority of cases, obtaining an -score of 89.7% and IoU of 81.4%. The developed code as well as the trained weights were made available in an open source project on GitHub, available at: https://github.com/pereira-gha/activefire, enabling other studies in the field.