The optimal segmentation scale identification using multispectral worldview-2 images

For the object-oriented image analysis, the quality of segmentation has a direct effect on classification accuracy. Segmentation scale parameter controls the sizes of image objects (or segments) and is the most important factor of the multi-scale segmentation quality of image segmentation. Therefore, identifying the optimal segmentation scale is very crucial for object-oriented land use classification and land use mapping. In this paper, a series of automated segmentations were firstly produced at a range of scales in eCognition Developer 8.0. Then, the unsupervised evaluation methods, which involve computing weighted variance, global Moran's I, Mean different to neighbor objects (Mean Diff.) and Ratio of Mean Diff. to Standard Deviation of image segmentations, were used to determinate the optimal segmentation scale in the multispectral World View-2 images (2 m) of a rural area. Meanwhile, the minimum mapping unit and mean object size of land use mapping were also analyzed and considered for the determination of the optimal segmentations. Finally, the comparison and analysis of experiment results show that the unsupervised methods, especially when combining intra-segment homogeneity and inter-segment heterogeneity, can obtain the optimal segmentation for object-oriented land use classification. The optimal segmentations derived from other methods probably contain some under-segmentation, but they are suitable to assist automated delineation of high spatial resolution images.