Automatic recognition of megalithic objects in areas of interest in satellite imagery

• Autor(a) principal: Caçador, David Galvão Chambel
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10071/21860
• topic: Multispectral images; Panchromatic images; Hyperspectral images; Circular hough transform; Spectral indices; Supervised machine learning; Imagens multiespectrais; Imagens pancromáticas; Imagens hiperespectrais; Transformação de hough circular; Índices espectrais; Aprendizagem automática

The integration of airborne and satellite imagery with Computer Vision and Machine Learning methodologies provided the ability of covering ample ground and enabled the detection of new terrain features, providing means to remotely locate, monitor and protect from destruction sites of cultural heritage. This work seeks to fuse spectral information obtainable from multispectral or hyperspectral images with spatial information derived from panchromatic images, provided by ESA and obtained from its satellites, to implement a high-performance automatic detection system capable of detecting buried or covered by vegetation dolmens. Separate methods were implemented, where for hyperspectral images a system was attempted based on the dolmens respective spectral material signature, and for panchromatic and multispectral images a system that extracted spectral indices, fused all into one and applied circle detection to identify dolmen locations, eliminating false positives through supervised machine learning. The hyperspectral images could not be used for the creation of a dolmens’ material signature, and by extension cannot automatically delineate regions of high likelihood of dolmen presence in images, due to the size of each pixel being of much higher dimensions than the known dolmens and the dolmens surrounding environments being too similar. The system created through the fusion of panchromatic and multispectral images proved capable of detecting dolmen locations, while simultaneously proving that part of the existent defined ontology of the dolmen (their insertion near water sources) can be used to delineate areas of high probability of dolmen presence, and that using supervised learning methods can enable the elimination of around 87.2% of false positives.