Detection of deforestation using remote sensing time series analysis

• Ano de defesa: 2020
• Autor(a) principal: Alber Hamersson Sánchez Ipia
• Orientador(a): Gilberto Camara Neto, Pedro Ribeiro de Andrade Neto
• Banca de defesa: Tiago Garcia de Senna Carneiro, Alexandre Camargo Coutinho
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Instituto Nacional de Pesquisas Espaciais (INPE)
• Programa de Pós-Graduação: Programa de Pós-Graduação do INPE em Ciência do Sistema Terrestre
• País: BR
• Link de acesso: http://urlib.net/sid.inpe.br/mtc-m21c/2020/06.30.14.25

The Amazon rainforest plays an important role in the global carbon and water cycles, having direct influence on Earths atmosphere and it suffers the consequences of the current climate crisis. Deforestation monitoring systems are a source of information on the forest condition for the scientific community, policy makers, and the general public. In this thesis, we identified three areas on which such systems could be improved: data processing, information extraction, and information distribution. Processing data of Earth observation satellites is subject to atmospheric noise. In particular, clouds obstruct the surveying of the Amazon rainforest. They introduce discontinuities on the the spatial and temporal patterns, which reduce the ability of analyst to extract information about features on the surface and reducing the reliability of the information obtained. Any information on Earths surface in our particular case, information on Land Use and Land Cover change increases its value through sharing, validation, and reuse in broader communities. Regarding data processing, we tested several cloud detection algorithms on Sentinel-2 imagery and we found that Fmask4 provides the best performance under frequent cloud coverage. With this knowledge, we proceed to extract deforestation information using time series of the Landsat 8 and Sentinel-2 satellites, applying machine learning techniques of Deep Learning and Random Forest, respectively. We obtained the best results by using time series of Sentinel-2 images processed with Random Forest. Finally, we demonstrated the best way to provide scientists with access to massive amounts or Earth observation data and processing tools is through collaborative analysis environments offered through Internet, such as Jupyter notebooks.