Medidas de acurácia baseada em objeto: análise metodológica em relação à validação baseada em pixel

• Ano de defesa: 2016
• Autor(a) principal: Prado, Daniel Fernando Costa do
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Lavras; Programa de Pós-Graduação em Engenharia Florestal; UFLA; brasil; Departamento de Ciências Florestais
• Palavras-chave em Português: Mapas – Projeção; Análise orientada a objeto; Análise baseada em pixel; Sensoriamento remoto; Map projection; Object-based image analysis; Pixel-based image analysis; Remote sensing; Solos Florestais; Sensoriamento Remoto
• Link de acesso: http://repositorio.ufla.br/jspui/handle/1/12479

The generation of thematic maps using images of high spatial resolution, applied to the new methodologies of analysis based on geographic object (GEOBIA), promotes a significant gain in the quality of land cover information. In addition, the maps generated from this approach bring great advantages over pixel-based analysis, especially in relation to the accuracy of the final map. The accuracy indexes used for the evaluation of these cartographic products are still obtained from the traditional validation, based on pixel. With the absence of a validation that represents the geographic object, new methodologies of object-based accuracy have been developed, such as the STEP similarity matrix. Therefore, the objective of this study was to compare an object-based validation methodology, in relation to the traditional pixel-based methodology, to evaluate the thematic land cover mapping. The study area includes part of the Lavras, Perdões, Ijaci, Bom Sucesso, Itumirim, Itutinga and Ibituruna counties, in the Minas Gerais State. The land cover map was generated from the RapidEye Image, acquired on 06/30/2011, with 5 meters of spatial resolution. The classification overall accuracy in relation to the traditional confusion matrix was 91.7% and 0.83 for the Kappa index. The reference objects used in object-based validation were vectored based on the targets that were contemplated by points, based on the same RapidEye image. The methodology used integrates four similarity measures (shape, theme, border and position) that result in three error matrices (individual and aggregated by thematic class and aggregated by weighted area). A global accuracy was obtained for the thematic similarity (82.25%) and for the similarity measures of position (89.93%), edge (98.47%) and shape (90.86%). The accuracy results of the producer and the traditional validation user, based on pixel, presented sub and overestimations among the classes mapped in relation to the similarity object-based accuracy, mainly between vegetation classes and anthropic area. The applied methodology is efficient in the evaluation of the objects and present great gains regarding the traditional validation, as much in relation to the thematic as geometric analysis.