Improved accuracy assessment indices for object-based high resolution remotely sensed imagery classification

High-resolution remote sensing images can capture detailed geometrical and shape properties. Traditional classification accuracy assessments with overall accuracy or kappa coefficient based on pixels, cannot exhibit the geometrical properties of the objects that are present on the ground. Evaluation of object oriented classified maps based on geometrical and border information can provide more accurate results. In this paper, we introduced and improved some object-based indices to evaluate the classification accuracy of the thematic maps obtained by high-resolution images. The indices depend on the geometry features of each object of the thematic map based on geometric error, including over segmentation, under segmentation, edge location, fragmentation error and shape error. Experiments conducted on Quickbird image in Fuzhou city show, compared to the traditional pixel-based accuracy assessment, our improved indices can provide more an accurately and quantitatively accurate evaluation of each land cover class, and can conduct more effectively for users to choose the best classification map.