Comparison Among Six Discrepancy Measures For Segmentation Evaluation In Object Based Image Analysis

Image segmentation is fundamental and essential for Object-Based Image Analysis because it will determine the upper accuracy limit of classification. An ideal segmentation will lead to a more efficient image analysis, so it is necessary to evaluate the segmentation. Discrepancy measures for remote sensing segmentation were be researched in this thesis. Discrepancy measure is an objective empirical method which evaluates the segmentation of some a combination of parameters by discrepancy between reference polygons and corresponding polygons. Discrepancy includes geometrical and arithmetical, a segmentation is regarded as the ideal one if reference polygons and corresponding polygons have same area and they are one-to-one correspondence.In order to compare the existing 6 discrepancy measures and explorer their properties fully, they are used to evaluate six segmentation datasets created based on multispectral and pan-sharpening multispectral images of Quickbird, WorldView II and ALOS. The conclusions drew from evaluate results are as follow:(1) ED2 and ED3-modified has best applicability, while ED2 is most sensitive.(2) Measures are correlation, ED2, ED3-modified and SEI have significant positive correlation, ED1 and F-measure have significant negative correlation.(3) For a credible user accuracy, the reference data with equivalent mean area should be used in the evaluation of segmentation and the accuracy assessment of optimal segmentation.(4) A widely accepted discrepancy measure should have veracity and sensitivity in any case.