Assessing and improving land use and cover samples using satellite image time series

• Autor(a) principal: Lorena Alves dos Santos
• Data de Publicação: 2021
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações do INPE
• Texto Completo: http://urlib.net/sid.inpe.br/mtc-m21c/2021/02.19.18.01

Land use and cover changes (LUCC) have caused a major impact on tropical ecosystems, increasing global greenhouse gas emissions and reducing the planets biodiversity. Remote sensing and digital image processing are powerful tools to measure and monitor LUCC effectively. Nowadays, with many Earth observation satellite images freely available, image time series analysis brings new opportunities and challenges for LUCC mapping over large areas. The use of remote sensing image time series analysis and machine learning to produce LUCC information has greatly increased. Machine learning using supervised techniques require a training step using land use and cover samples labeled a priori. For this reason, it is necessary highquality samples to avoid negative effects in classification performance. Due to the recent availability of open Earth observation data, methods using satellite image time series still are a gap in the literature. This thesis contributes to Earth observation field proposing two methods to assess the samples quality and reduce the noise in the land use and cover reference datasets. The main idea is to identify mislabeled samples, data with low discrimination when mixed with other classes, and explore the samples spatiotemporal variability using satellite image time series. The first method is based on unsupervised neural networks, the self-organizing map (SOM), and Bayesian Inference. It provides measures to identify mislabeled samples and assess the reliability of the samples. In the second method, the hierarchical clustering is combined with SOM to generate subgroups to identify spatiotemporal patterns to explore the samples intra-class variability. Both methods use satellite image time series. It allows the Earth observation scientists to understand the samples behavior over time, contributing to noise reduction in land use and cover reference databases. These methods were applied in different case studies using samples in the Cerrado biome in Brazil.The results indicated that the methods are efficient to reduce class noise and to assess the spatio-temporal variation of satellite image time series training samples.