The Study Of Object-oriented Remote Sensing Image Multi-scale Segmentation's Optimal Parameters

In the field of remote sensing image analysis,object-oriented image analysis method has replaced the image analysis method based on pixels.Using object-oriented image analysis method,we need to segment image into several meaningful objects.Those objects would have geometry information,texture information and spatial information except spectral information.Remote sensing image has the characteristics of multi-scale,using a single scale segmentation method could not obtain good result,so multi-scale segmentation gradually caused the attention of more and more experts and scholars.Some multi-scale segmentation theories and methods have been proposed,some of them are actually used and achieve good results.However the multi-scale segmentation has a strong parameters dependence,in this paper we solve the problem of the determination of multi-scale segmentation parameters.Multi-scale segmentation commonly uses region merging algorithm which is based on principle of minimum heterogeneity.Its common parameters include scale parameter and heterogeneity parameters.By far the experts and scholars focus on the determination of the optimal scale parameter,but rarely study heterogeneity parameters such as spectral feature and shape feature weighting factor,in the process of choosing the optimal segmentation parameters they mainly focus on the optimal parameters of specific types of feature,but rarely study the global optimal segmentation parameter selection.In this paper we will calculate the optimal value of all the main parameters in multi-scale segmentation in the case of fractal net evolution approach,mainly there are scale parameter,spectral feature weighting factor,shape feature weighting factor,image tightness weighting factor and smoothness weighting factor.This paper would deal with two cases.In first case we would calculate the optimal segmentation parameters of some kinds of feature,which include building,road,vegetation and water features.In the other case we would calculate the global optimal segmentation parameters.For the former case,we analyze the result of segmentation,then calculate area of specific surface based on surface features,finally select the parameters corresponding largest area as the most optimal parameters,which proved behaves well in experiment.We conduct result by control contrast method and orthogonal experiment method,then compare these result and find that orthogonal experiment method behaves better.For the latter case,we analyze the result of segmentation,then calculate normalized index of maximum and minimum area and area distribution entropy to determine the most optimal parameters.We compare the result with the result of mean standard deviation method,experiments show that the method of this paper achieve better result.