Métodos de ajuste para fusão espaço-temporal de imagens de satélite

• Ano de defesa: 2024
• Autor(a) principal: Caon, Ivã Luis
• Orientador(a): Mercante , Erivelto
• Banca de defesa: Mercante, Erivelto , Mendes, Isaque de Souza , Johann, Jerry Adriani
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Estadual do Oeste do Paraná; Cascavel
• Programa de Pós-Graduação: Programa de Pós-Graduação em Engenharia Agrícola
• Departamento: Centro de Ciências Exatas e Tecnológicas
• País: Brasil
• Palavras-chave em Português: Sensoriamento remoto; Computação em nuvem; Classificação de uso e cobertura da terra.
• Palavras-chave em Inglês: Remote sensing; Cloud computing; Land use and land cover classification
• Área do conhecimento CNPq: CIENCIAS AGRARIAS::ENGENHARIA AGRICOLA
• Link de acesso: https://tede.unioeste.br/handle/tede/7500

Remote sensing enables the acquisition of data over vast geographic areas, making it essential for characterizing land use and land cover. However, the availability of highresolution data is often limited by factors such as adverse atmospheric conditions. Spatiotemporal fusion emerges as a practical solution by combining the advantages of different sensors to produce high-resolution spatial and temporal data. Since the development of the Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM) in 2006, numerous fusion methods have been proposed, ranging from pairwise value compilation techniques to the use of deep learning for establishing relationships between sensors. Although efficient, these techniques are computationally expensive and require advanced technical expertise. Additionally, they face challenges in constructing accurate patterns and detecting abrupt changes in land cover, such as wildfires and floods. As a viable alternative for remote sensing data processing, Google Earth Engine (GEE) offers high-performance computing resources on an accessible and practical platform. It facilitates the dissemination of results to other researchers, thereby promoting scientific collaboration. In this regard, the results obtained indicate that spatiotemporal fusion using the 4th-degree Least Squares model in the OLI/MODIS system achieved the best performance. While it did not yield the highest coefficient of determination (r²), it achieved the best overall accuracy. Statistical analysis confirmed that this model performs stochastically equal to higher-degree models but with reduced complexity, making it the most efficient choice. On the other hand, the MSI/MODIS system, even with fusion using the 1st-degree Talwar model, did not show significant gains compared to MSI without fusion. This outcome is likely due to the lower compatibility between the spectral resolutions of the sensors involved, which limited the expected benefits. Spatiotemporal fusion proved promising for the OLI/MODIS system. However, further studies are necessary in order to optimize the fusion methodology for the MSI sensor, considering adjustments to better leverage the potential of the involved sensors or exploring the use of a different sensor in place of MODIS. In conclusion, despite persistent challenges, spatiotemporal fusion based on regression techniques represents a promising and innovative approach to enhancing the resolution and applicability of remote sensing data, with vast possibilities for application across various fields of knowledge.