Remote Sensing Image Segmentation

The information extraction of high-resolution remote-sensing images is of greatimportance for mine monitoring, and the image segmentation is a critical step for theinformation extraction of the high-resolution remote-sensing images. For the current popularmulti-scale segmentation technology, the fixed scale of the data source has been avoided. Thetechnology combines the advantages of both the coarse-scale segmentation and the fine-scalesegmentation, in which the former segmentation method can be used to shield the spatialnoise caused by numbers of physical factors from the latter method, and the advantages ofmuch abstract information contained in the former can be well employed. In addition, thelatter method can be used to conduct the thinning of the image segmented by the formermethod. However, currently the application of the segmentation method of high-resolutionremote-sensing images ofrare earth mines is not sufficient; the multi-scale segmentationmethod is not targeted for any specific high-resolution remote-sensing images; for the imagesegmentation results, at present no uniform precision evaluation system is established so thatthe better segmentation results cannot be identified affirmatively.On the basis of the three above-mentioned problems, taking the Ganzhou City DingnanCounty with rich rare earth mine as the research area, in consideration of the features of thesurface objects of the mining area, the image data of Resource No.1satellite-02C (ZY-102C),Quick Bird and Landsat8will be selected for experiments, and the main tasks andachievements of the research are as follows:1. After the referencing of the technical methods from various domestic and foreigndocuments and the comparison of their advantages and disadvantages, the FNNA(Fractal Net Evolution Approach) algorithm is selected as the segmentationalgorithm and through the algorithm, the subjective factors can be effectivelyavoided. In the paper, the method of automatically determining the threshold valueis provided, andthe determination of the algorithm and the threshold value isachieved.2. Article compares the advantages and disadvantages of the maximum area method,the average local variance method and the neighborhood mean absolute difference variance(RMAS), through the experiment, RMAS was chosen as a way todetermine the features’ optimal scale. The reason is that the experimental results ofthree methods the tank’s optimal scale respectively is10,20and15. It can beobtained through comparing that tank’s optimal scale is15which match theoptimal scale15with the best visual image by FNEA.3. Determine the target objects: tanks, houses, vegetation and bare ground, preprocessimages, set segmentation parameters, do segmentation experiments to imageZY-1-HR, ZY-1-P/MS, QB and L8with scale5-60by step5.In the segmentationresult of images ZY-1-HR and QB: tank’s optimal scale is15, house’s optimalscale respectively is20and30, vegetation and bare ground optimal scale are45and50. Due to lower resolution, only houses, vegetation and bare ground can beidentified in image ZY-1-P/MS, and their optimal scale is respectively25,40;however, only vegetation and bare ground can be identified in image L8, and theiroptimal scale is35. The optimal segmentation scales are different which of varioustargets objects in a particular image or same object in different resolutionimages.For the same object in the same area, the optimal is on the decline withlower resolution. Each feature has its suitable segmentation methodand suitableresolution for segmentation,the optimal segmentation scale for the surface objectsis relative value.4. An objective and general evaluation index is established, and on the basis of theindex, the segmentation quality value and the evaluation relationship between thesegmentation scales are established, then the segmentation scale diagram related tovarious surface objects and the resulting segmentation quality values is comparedwith the RMAS diagram witch get that the maximum value of segmentationqualityis same as the maximum of RMAS,it verify that the optimal segmentation scalethrough the RMAS method is correct.