Study On Object-oriented Remote Sensing Image Segmentation And Classification Methods

The emerge of high-resolution remote sensing images has greatly promoted theapplication of remote sensing technology in various fields. Hot topics of currentresearch is how to extract the target information from the high-resolution images. Inthis paper, high-resolution images of multi-source data source was used to studied theobject-oriented remote sensing image classification techniques. Firstly, the principlesand implementation process of image segmentation algorithm was studied, fromprimary segmentation and region merge two aspects,to improve the efficiency andaccuracy of image segmentation. Secondly, the characteristics quantitative methods ofdifferent objects and the effect of different characteristics to the image classificationresults were researched. Finally, support vector Machine (SVM) classification methodand the nearest neighbor method was chosen to classify the image objects to achieve ahigh-resolution remote sensing image.（1）Image segmentation algorithm. In order to improve the efficiency of imagesegmentation, the primary segmentation algorithm uses only one pass of thewatershed segmentation algorithm. In order to improve the over-segmentation ofwatershed segmentation, marked a watershed segmentation method using dynamicthreshold segmentation of remote sensing images at the beginning, which greatlyimproved the early image segmentation. To further optimize the primary segmentimage object, the region adjacency graph (RAG) was applied to express therelationship between neighborhood adjacent regions, region merging criterion isbased on a combination of spectral shape segmentation results of the merger in earlyto get a different scale by setting different thresholds merger segmentation resultsunder optimal scale exploration of image segmentation.（2）analysis and combinations of image object characteristics. Image objectconsist of spectrum, shape and texture characteristics.Study and analyze therelationship between the different characteristics of the various types of surfacefeatures; according to different classification purposes, selecting appropriate combination of features, the spectrum from multiple angles, shapes and textures,such as refined classification results improve feature extraction accuracy.（3） Object-oriented high-resolution remote sensing image classification.SVM algorithm and the nearest neighbor classification algorithm were used fordifferent remote sensing image classification experiments to achieve theobject-oriented high-resolution remote sensing image classification. Then, analysisthe advantages and disadvantages of the two methods in the object-orientedclassification.Research shows that segmentation algorithm used in this paper to achieve theregional dynamics at the beginning of remote sensing image segmentation, to acertain extent the traditional watershed algorithm to improve the over-segmentationphenomenon; using object-oriented herein segmentation and classification algorithmsfor ZY-3and WorldView-2images primary segmentation, multi-scale object-orientedclassification and region merging, the experiment proved that the object-orientedinformation extraction method is feasible and exists a certain advantage.