A carregar...
Publicação
Segmentação avançada de Faixas de gestão de combustível
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 5.93 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
A constante evolução e investimento na aquisição de informação sobre o nosso planeta
através de satélites resultou numa quantidade enorme de dados que é humanamente
impossível de processar, analisar e extrair conclusões. Para solucionar este problema,
surgiram novas tecnologias e métodos para realizar essas tarefas, tirando proveito da
informação adquirida. Nesta dissertação serão estudadas algumas destas técnicas.
O Instituto de Conservação da Natureza e das Florestas (ICNF) definiu a implemen-
tação da rede primária de Faixas de Gestão de Combustível (FGC) como estratégia para
prevenção e combate a incêndios florestais. Nessas faixas são realizadas intervenções
para reduzir a vegetação nas mesmas, servindo como barreira à progressão do fogo. No
entanto, devido ao elevado tamanho da rede, é difícil de a gerir e supervisionar. Para
resolver este problema, foram criadas técnicas automáticas de deteção de manutenções
nas faixas usando imagens do Sentinel-2.
Nesta dissertação foram estudadas duas metodologias para a deteção de tratamentos
nas faixas. A primeira é uma metodologia baseada numa aprendizagem supervisionada e
a segunda baseia-se no teste estatístico Welch’s t-Test, metodologias de deteção com uma
abordagem ao nível do píxel e ambas têm um pré-processamento comum, o co-registo, que
consiste num alinhamento geométrico das imagens de satélite para garantir que o mesmo
píxel represente o mesmo local nas várias imagens. Foi dado maior ênfase à segunda
metodologia, identificando possíveis maneiras de melhorar o seu desempenho, através
da variação de parâmetros e a comparação dos resultados obtidos com um conjunto
de dados de validação. Por fim, para esta metodologia foi proposto e avaliado um pós-
processamento para reduzir o número de classificações isoladas.
The continuous evolution and investment in acquiring information about our planet through satellite technology has resulted in an overwhelming amount of data that is beyond human capability to process, analyze, and derive meaningful conclusions from. To tackle this challenge, new methods and technologies have emerged to perform these tasks, leveraging the collected information. In this dissertation will be studied some of these techniques. The Institute of Nature and Forest Conservation (ICNF) has established the imple- mentation of the primary network of Fuel Breaks (FGC) as a strategy for preventing and fighting forest fires. Interventions are carried out in these zones to reduce vegetation, and their primary objective is to inhibit the spread of fire. However, due to the vast size of the network, it is challenging to manage and supervise effectively. To solve this problem, automatic techniques have been developed for maintenance detection in the Fuel Breaks using Sentinel-2 images. In this dissertation, two methodologies were studied for detecting treatments in the Fuel Breaks. The first is a methodology based on supervised learning, while the second is based on the Welch’s t-Test statistical test. They are detection methodologies with a pixel- level approach and both have a common pre-processing, co-registration, which consists of a geometric alignment of satellite images to ensure that the same pixel represents the same location across multiple images. Greater attention was given to the second methodology, with the aim of identifying potential ways to enhance its performance by varying parameters and comparing the results obtained with a validation dataset. Finally, a post-processing was proposed and evaluated for this methodology to minimize the number of isolated classifications.
The continuous evolution and investment in acquiring information about our planet through satellite technology has resulted in an overwhelming amount of data that is beyond human capability to process, analyze, and derive meaningful conclusions from. To tackle this challenge, new methods and technologies have emerged to perform these tasks, leveraging the collected information. In this dissertation will be studied some of these techniques. The Institute of Nature and Forest Conservation (ICNF) has established the imple- mentation of the primary network of Fuel Breaks (FGC) as a strategy for preventing and fighting forest fires. Interventions are carried out in these zones to reduce vegetation, and their primary objective is to inhibit the spread of fire. However, due to the vast size of the network, it is challenging to manage and supervise effectively. To solve this problem, automatic techniques have been developed for maintenance detection in the Fuel Breaks using Sentinel-2 images. In this dissertation, two methodologies were studied for detecting treatments in the Fuel Breaks. The first is a methodology based on supervised learning, while the second is based on the Welch’s t-Test statistical test. They are detection methodologies with a pixel- level approach and both have a common pre-processing, co-registration, which consists of a geometric alignment of satellite images to ensure that the same pixel represents the same location across multiple images. Greater attention was given to the second methodology, with the aim of identifying potential ways to enhance its performance by varying parameters and comparing the results obtained with a validation dataset. Finally, a post-processing was proposed and evaluated for this methodology to minimize the number of isolated classifications.
Descrição
Palavras-chave
Deteção Remota Faixas de Gestão de Combustível Sentinel-2 Aprendizagem Supervisionada Welch’s t-Test Teste de Normalidade