Automating land cover change detection: a deep learning based approach to map deforested areas

• Ano de defesa: 2020
• Autor(a) principal: Raian Vargas Maretto
• Orientador(a): Leila Maria Garcia Fonseca, Thales Sehn Körting
• Banca de defesa: Rafael Duarte Coelho dos Santos, Rogério Galante Negri, Nathan Jacobs
• 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 Computação Aplicada
• País: BR
• Link de acesso: http://urlib.net/sid.inpe.br/mtc-m21c/2020/06.09.11.59

Accurate maps are an important tool for informing effective deforestation containment policies. The main existing mapping approaches to produce these maps are largely manual, requiring significant effort by trained experts. In recent years, Deep Learning (DL) have emerged becoming the state-of-the-art in Machine Learning and Pattern Recognition. Despite its effectiveness, the computational concepts behind these methods are very complex, as well as the computational platforms available to implement it. This complexity makes it difficult for a Remote Sensing analyst without a strong programming background to perform image analysis using those methods. Furthermore, despite DL have been successfully applied in many Remote Sensing studies, most of those have focused on the detection of very specific urban targets in high-resolution imagery, due to the high availability of reference and benchmark datasets with these characteristics. The lower number of studies on the application of DL to medium and low-resolution imagery and to another types of targets have been attributed, among other reasons, to the lack of reference and benchmark datasets for these types of images. Within this context, this thesis has three main contributions. First, we developed DeepGeo, a toolbox that provides modern DL algorithms for Remote Sensing image classification and analysis. DeepGeo focuses on providing easy-to-use and extensible methods, making it easier to those analysts without strong programming skills to use those DL methods. It is distributed as free and open source package and is available at https://github.com/rvmaretto/deepgeo. Second, we present the PRODES-Vision collection of dataset, a collection of reference dataset of deforested areas, based on PRODES deforestation maps, to train Deep Neural Networks, as well as a methodology to the generation of reference datasets based on thematic maps. We believe that these datasets would encourage the development of new methods for automatically map Land Use and Land Cover changes. And finally, we propose a fully automatic mapping approach based on spatio-temporal convolutional neural networks aiming to reduce the effort of mapping deforested areas. Furthermore, we propose two spatio-temporal variations of the U-Net architecture, which make it possible to incorporate both spatial and temporal contexts. Using a real-world dataset, we show that our method outperforms a traditional UNet architecture, achieving approximately 95% accuracy. We also demonstrate that our preprocessing protocol reduces the impact of noise in the training dataset. To demonstrate the scalability of our method, it was applied to map deforestation over the entire Pará State, achieving approximately 94% overall accuracy. And finally, to demonstrate its applicability to another areas, it was applied to a region of the Brazilian Cerrado, achieving approximately 91% overall accuracy.