Research On High-resolution Image Segmentation Method Based On Multi-criteria Histogram Threshold Integration And Region Merging

With the development of earth observation technology,high spatial resolution remote sensing images are more convenient to be obtained and more widely used,the pixel-based classification methods no longer meet the processing needs,and objectoriented method gradually becomes more important and plays an important role in high spatial resolution image analysis.Image segmentation is the most basic and key step in object-oriented method,Most image segmentation methods require manual parameter adjustment to obtain the optimal segmentation results,which are highly subjective and time-consuming.Traditional threshold segmentation,which is widely used in image segmentation but easy to produce a lot of broken spots in image,often leads to unsatisfactory segmentation and has a large extent limits of the segmentation in high-resolution images.Based on these two starting points,the paper proposes a segmentation method by integration multi-criteria histogram thresholding and region merging,and validates the method by supervised classification of typical land cover in two study areas of WorldView-2 orthophoto image.The main contents and conclusions are as follows:(1)Construction of multi-criteria histogram threshold integration segmentation methodThe method included segmentation ensemble of single image feature and that of multi-image features,and the latter was based on the former.12 histogram thresholding segmentation methods were selected to segment the image features.Firstly,the different threshold segmentation results of single image features were overlaid to produce the homogeneous region,and the connected domain analysis and single pixel processing were carried out to obtain one segmentation integration image.Then,the segmentation images of other image features were overlaid each other.After the connected domain analysis,one segmentation image was obtained.(2)Region merging method applied to multi-criteria histogram threshold integration segmentationRegion merging was performed by a two-stage strategy,and the mean square error was used as the merging criterion in both stages.The first merging removed the isolated pixels in the initial segmentation image and reduced the merging time in the second stage,the merging was terminated when there was no region with pixel number 1.The second merging used the adaptive method to determine the optimal scale,based on the stepwise evolution analysis framework.The methods used included stepwise evolution of the area-weighted variance in the region merging process and the univariate linear regression.(3)Application validation of multi criteria histogram threshold integration segmentation method in land cover classificationBased on the supervised classification of typical land cover in two study areas of WorldView-2 image,in which SVM classifier was used,proposed method was validated.The results show that the overall classification accuracy and Kappa coefficient based on the proposed method are 89.32%,89.28% and 0.81,0.79,respectively.Compared with overall classification accuracy(OA)of pixel-based,segmentation of simple linear iterative clustering algorithm,and segmentation of multiresolution,OA based on segmentation object from proposed method was improved by 2.01%,8.15%,3.18% in the first area and 9.17%,12.1%,2.83% in the second area,respectively,and the kappa coefficient was improved by 0.03,0.14,0.06 and 0.13,0.2,0.05,respectively.The proposed method can segment different sizes of ground objects by a single scale,and get the best for the small area of vegetation distribution.In conclusion,the segmentation method based on multi criteria histogram threshold integration shows good application value in land cover classification,and the overall accuracy is better than that of pixel-based and segmentations from other methods used for comparison.The results of this paper provide reference and support for object-oriented segmentation in high spatial resolution imagery.