Research On The Segmentation Optimization In Object-based Image Analysis For Extracting Typical Land Cover From High Resolution Images

In object-based image analysis(OBIA),segmentation optimization is the first and indispensable step which has direct effects on the quality of the subsequent classification.Current methods of parameter optimization cope with many problems,such as low precision of image segmentation,poor operability,limited application,and so on.There is still urgent need to develop a segmentation optimization method with respect to high accuracy,high simplicity,and high universality.By using the Fractal Net Evolution Approach(FNEA)for image segmentation,this paper proposes a novel segmentation optimization function(SOF)based on an analysis of relations between the internal homogeneity,neighborhood heterogeneity of typical image objects and scale parameter.By comparison with discrepancy indicator ED2,which performs evident advantages on segmentation parameter optimization,this study draws following results:(1)The internal homogeneity of the matching objects of the typical land covers shows a decreasing trend with the increase of the segmentation parameters,while the neighborhood heterogeneity increases with the increase of the segmentation parameters.The SOF value shows a trend of U shape with the increase of the scale parameter,while the optimal segmentation parameter appears at the bottom of the U shape curve.The characteristics of "U" shape curve depend on the variation of internal homogeneity and neighborhood heterogeneity with the scale parameter.The left side of the "U" type curve mainly depends on the variation characteristics of neighborhood heterogeneity.The bottom part of the curve depends on the synthetic effects of both inter nal homogeneity and neighborhood heterogeneity.The right side mainly depends on the internal homogeneity change characteristics.(2)The variation of internal homogeneity,neighborhood heterogeneity and SOF of the entire image is the same as typical land cover types.The SOF can not only be used to in segmentation parameter selection for typical land covers,but also for the entire scene of image.(3)The results of SOF and PSE–NSR–ED2 system for segmentation optimization are close each other.In addition,the optimum segmentation parameter for SOF is a single scale,and the optimal segmentation parameter of ED2 may be a scale range(4)The result of segmentation optimization based on ED2 assessment is prone to under-segmentation slightly,while SOF is prone to over-segmentation slightly.When the studied objects distribute in a large area in space,and the boundary between the objects is not evident,SOF and ED2 tend to under-segmentation,while the data based on SOF assessment is more obvious.(5)Segmentation optimization effect of SOF and ED2 are good,but SOF is slightly better than ED2.The results of typical land covers information extraction in the study area show that the accuracy of information extraction on SOF and ED2 is more than 70%,which indic ates that the parameters of SOF and ED2 are good.Among them,there are 14 kinds of typical land covers,the accuracy of information extraction on SOF optimal segmentation parameters is better than that of ED2.There are 3 kinds of typical land covers,the accuracy of information extraction on SOF and ED2 optimal segmentation parameters are the same.There are 5 kinds of typical land covers,the accuracy of information extraction on ED2 optimal segmentation parameters is better than that of SOF.Therefore,the optimal effect of SOF segmentation parameters is better than ED2.