Study On Selecting The Optimal Parameter For Remotely Sensed Image Segmentation

Object-based image analysis method is an effective method of remote sensing image classification. This method is mainly divided into two steps:image segments and image objects classification. The process of image segmentation is use multi-scale segmentation methods to split remote sensing image into many segmentation objects, and object-based classification is found classification rule-sets with many features of image segmentation objects, and use the rule-sets to classify the image segmentation objects into the classes we wanted. In the segmentation stage, the multi-scale segmentation techniques can give us different segmentation results of difference levels, while different land-cover types requires segmentation results with different levels, so we should choose optimal segmentation parameter for a particular land-cover type. The image segmentation is the foundation and the key of the object-based image analysis, the accuracy of image segmentation determines the accuracy of classification directly, the research with the optimal parameter of image segmentation is particularly important.Based on deeply analysis of remote sensing image segmentation algorithms, a study about optimal parameter of image segmentation has been done. The main contains and results are as follow:(1) The paper analysis some typical algorithms about remote sensing image multi-scale segmentation:one is region merging algorithm based on heterogeneity minimum principle; another is the algorithm based on watershed transform about gradient map of image. The scale parameter of multi-scale segmentation represents the fragmentation level of image segmentation, the multi-scale segment results have a hierarchical structure, and the vector boundary of multi-scale segment results is inherited from parent. An experiment analysis segment results from watershed transform, and shows that the vector boundary of segment results is not strictly inherited.(2) A discrepancy method based on regional discrepancy measures is an emphasis study, which can choose optimal parameter of image segmentation. Analysis of the method shows that a problem about the segment results based on watershed transform; the arithmetic discrepancy changed greatly because of watershed transform was sensitivity about the boundary in remote sensing image. So the paper improve this discrepancy method based on regional discrepancy measures, and propose another two normalized indices to measure the geometric difference and arithmetic difference, Under-segmentation of Area Difference Index (UADI) and Number Difference Index (NDI). It is also calculate the Euclidean distance of UADI and NDI, and find the optional parameters of image segmentation.(3) The technique about program design based on ENVI/IDL and ArcGIS was used to calculate the optional parameters of the two methods above. The technique enforced by the segmentation algorithms about ENVI EX and the functions about space analysis of ArcGIS.(4) The study area is the city of Zhongwei in Ningxia Autonomous Region. The experiment data is multi-spectral band of Landsat TM images and multi-spectral band of ALOS images. The experiment calculates the optimal parameter about pond, residential areas, farmland and artificial forest. Then calculate Average Spectrum Heterogeneity about corresponding objects of each land-cover types, to evaluate these two methods. The experiment shows that the improved method has the most effectiveness.