Segmentação não supervisionada de imagens de sensoriamento remoto por minimização da entropia cruzada

• Autor(a) principal: Santana, Eduardo Freire
• Data de Publicação: 2014
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da UFPB
• Texto Completo: https://repositorio.ufpb.br/jspui/handle/tede/6134
• topic: Segmentação Não Supervisionada de Imagens; Modelos Estatísticos; Informação; Entropia; Sensoriamento Remoto; Unsupervised Segmentation; Statistical Models; Information; Entropy; Remote Sensing; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO

Remote sensing is one of the fastest growing technologies of late twentieth and early twenty-first century. The most common use of this term is related to the optical sensing of Earth's surface through satellites. In remote sensing, image segmentation is a process often used to aid in landscape change detection and land use classification. This study aims the research and development of a new method for unsupervised segmentation of remote sensing images by minimizing the cross entropy between the probability distribution of the image and some statistical model. Images used for tests were captured by the Thematic Mapper sensor on Landsat 5 satellite. The proposed algorithm takes an initial segmentation and progresses iteratively, trying to improve the statistical model and reduce the cross entropy with respect to previous iterations. Results indicate that the cross entropy minimization is related to a consistent image segmentation. Two approaches were developed, one by performing a per-pixel classification and the other by classifying regions obtained by the Watershed transform. In per-pixel approach, the average agreement between the classifier and the thematic image used as ground truth was 88.75% for fifteen selected images and 91.81% for four small regions that represent details of land use transitions, such as vegetation, rivers, pastures and exposed soil. In region approach, the average agreement was 87.33% for images and 91.81% for details. The ground truth for image details was manually created by an expert.