Land cover automatic classification using deep learning techniques applied to satellite imagery

• Autor(a) principal: Santos, Sérgio Filipe Paiva da Silva Gonçalves dos
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10071/31072
• topic: Multispectral imaging; Rede neuronal convolucional - -- Convolutional neural network (CNN or ConvNet); Machine learning; Land use land cover classification; Sentinel-2; Imagens multispetrais; Classificação da superfície terrestre

In this era of population growth and rapid urbanization, effective and sustainable urban development is a very important factor. In this context, Machine Learning (ML) can play a leading role in helping with these tasks; it makes possible the treatment of remote sensing images in a shorter time frame. This thesis focuses on the development and use of Convolutional Neuronal Network (CNN)s to handle multispectral images. The principal goal is to evaluate the performance metrics and computational complexity of a CNN-based land cover classification approach. And to try and assess if the results achieved are better or worse than the architectures currently implemented by the Direção Geral do Território (DGT). In this order, it was first necessary to understand the provided data and all its inherent characteristics. This data was then preprocessed, and the architecture was defined. The results show that CNNs present a promising alternative in this context to the implemented methods for land cover classification. Despite the promise it provides, it also highlights the difficulties faced and how the work can be improved, specifically concerning the lack of labeled data. The existence of these difficulties presents opportunities for further development of this work. As an overview of this dissertation, it is possible to say that the investigation into the feasibility of using CNNs for land cover classification provided positive results. There is, however, as would be expected, room for improvement, especially in what concerns the pre-processing of data.