An evaluation approach for segmentation results of high-resolution remote sensing images based on the degree distribution of land cover networks

Abstract

Identifying the degree distribution of land cover networks is helpful to find analytical methods for characterizing complex land cover, including segmentation techniques of remote sensing images of land cover. After segmentation, we can obtain the geographical objects and corresponding relationships. In order to evaluate the segmentation results, we introduce the concept of land cover network and present an analysis method based on statistics of its degree distribution. Considering the object-oriented segmentation and objects merge-based spectral difference segmentation, we construct the land cover networks for different segmentation scales and spatial resolutions under these two segmentation strategies, and study the degree distribution of each land cover network. Experimental results indicate that, for the object-oriented segmentation, the degree distributions of land cover networks follow approximately a Poisson distribution, regardless of the segmentation scales and spatial resolutions. For the objects-merge method based on spectral difference segmentation, degree distributions exhibit heavy tails. Compared with all the segmentation results, the pattern spots after objects-merge better retain the integrity of geographical features and the land cover network can reflect more accurately the topological properties of real land cover when the threshold of objects merge is suitable. This study shows that we can evaluate the reliability of segmentation results objectively by analyzing the degree distribution pattern of land cover networks.