Accuracy evaluating method for object-based segmentation of high resolution remote sensing image

Abstract

Based on the principles of supervised evaluation method, the paper puts forward three evaluation indices of the high resolution remote sensing image segmentation: precision, recall and relative similarity, and brings forward the method of precision evaluation of remote sensing image segmentation. With respect to the reference object matching problem of the supervised evaluation method, the paper proposes the matching method of bidirectional local optimal object. Validation of the proposed evaluation indices is carried out using GF-1 high resolution remote sensing image in Huainan city, Anhui province, China. The results show that the proposed evaluation indices can reflect very well the quality of the segmentation results and are consistent with the real distribution of ground landcover segmentation, and also provide the basic reference to parameter setting for the image segmentation algorithm and optimal scale selection for the multi-scale segmentation.