Detecção de incêndios em imagens de sensoriamento remoto: um estudo integrando os satélites Landsat-8, Sentinel-2 e inteligência artificial

• Ano de defesa: 2021
• Autor(a) principal: Pereira, Gabriel Henrique de Almeida
• 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: Incêndios; Inteligência artificial; Sensoriamento remoto; Aprendizado de máquinas; Imagens de sensoriamento remoto; Fires; Artificial intelligence; Remote sensing; Machine learning; Remote-sensing images; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO; Ciência da Computação
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/27073

Active Fire detection in satellite imagery is of great importance for the management of environmental conservation policies, and environmental monitoring, supporting decision-making, law enforcement and even preventing financial or human losses. Active Fire detection in satellite imagery is a well-established field, with many techniques being proposed over the years, often based on pixel or region-level comparisons involving sensor-specific and neighborhood statistics. For some satellites this problem is already well resolved. For others, some changes in the established methods are enough to create new algorithms. And for some other satellites this is still an open issue. Landsat-8, for example, which already has methods consolidated in the fire detection literature, while for Sentinel-2 the methods are still incipient. However, given the similarities between some bands of these satellites, especially in the bands sensitive to the presence of fire, it would be possible to use these satellites integrated for the active fire detection purpose. However, variations in sensor sensitivity may mean that methods developed for Landsat-8 are not directly applicable to Sentinel-2 images, especially the methods based on thresholds. Artificial intelligence techniques, which have enjoyed enormous success in analyzing remote sensing data, can provide solutions for these situations. This work thus addresses the problem of detecting active fire using artificial intelligence techniques on images from these two satellites (Landsat-8 and Sentinel-2). Both machine learning and deep learning methods are tested in this work to identify which approaches present better behavior for fire detection in these satellite images. This work also pays attention to the training process of these artificial intelligence models, either with some tests of the models’ parameters, variations in band combinations and training sample proportions. The experiments — for a dataset developed for this work, on a global scale, has been made publicly available — showed that the artificial intelligence methods, both machine learning and deep learning, were able to make active fire detections in both Landsat-8 images (with some models with -Score up to above 0.9), as well in Sentinel-2 images (with -Score up to 0.8). Finally, it is noteworthy that the artificial intelligence methods, more than just being able to generate fire detections in the Landsat-8 and Sentinel-2 satellite images, were able to be trained with Landsat-8 information and make inferences on Sentinel-2 images.