Comparação entre abordagens orientada a objetos geográficos e pixel a pixel para classificação supervisionada de imagens MSI / Sentinel-2

Comparison between approaches to geographic objects and per pixel for supervised classification of MSI / Sentinel-2 imagery

Abstract

The traditional approaches of sorting of the orbital images per pixel consider only the spectral attributes of the image. The classifiers based on geographic objects (GEOBIA), consider besides spectral attributes, the shape, the size, the texture, and the occurrence. In this context, the work aims to do a comparison between GEOBIA and per Pixel methods to the supervised classification of land use and coverage with high resolution imagery. The study has been carried out on a 72 km² cutout of the 22JFP scene of the Sentinel-2 satellite. In the classification by GEOBIA, image segmentation, attribute extraction and supervised classification with the C4.5 algorithm were performed. In the pixel approach, the maximum likelihood algorithm (MAXVER) was used. The classification by GEOBIA showed agreement rates (Global Accuracy, Kappa and conditional Kappa) higher than those of the per Pixel classification. Regarding the individual accuracy of the classes, both presented satisfactory results.