Comparative Study On Object-oriented Classification Technique With The World-view2 High-resolution Image

With the development of remote sensing technology, especially the development of the sensor, the resolution of the remote sensing image is greatly improved, make it more and more wide that the remote sensing images used in land use coverage, as the original develop from the Landsat7 spatial resolution of 30 meters to the current development of submitter high resolution images.The high spatial resolution imagenot only has thespatial-detail informationin detail,but alsohas the richband information. However, the traditional pixel based classification Technique not only can effectively use the spatial texture information, shape information and spatial topological relation of high resolution image, but also make the result of classification with the phenomenon of salt and pepper, resulting in a large number of invalid broken patches. Therefore, to obtain high spatial resolution images in the use of land use and land cover information, we always use artificial visual interpretation. But this method is so high cost and need a lot of human and material resources to complete. From the above, the traditional image interpretation method has been unable to meet the needs of high spatial resolution image information extraction work.World-View2 as a high resolution commercial satellite, which has high resolution and rich band information, is widely used in the actual production of life. However, in the practical application, in order to ensure the accuracy of classification, we often use manual visual interpretation method. To solve the above problems and make the extraction of high resolution World-View2 image information more efficient and accurate, this study which is based on the World-View2 band informationevaluate and analyze the object oriented classification methods that include degree of membership function method, k-nearest neighbor, decision tree classification, by using multi-scale segmentation method and the comparison of translation results and artificial visual solutions.This study takes the World-View2 image of Ludian County as the research object to get a resolution of 0.46 M multi-spectral image by preprocessing the World-View2 image data and doing World-View2 image band fusion. Then we do the research of image multi-scale parameter segmentation, so that we can obtainoptimal parameters for each classification category. At last, this study which is based on the World-View2 band information evaluate and analyze the object oriented classification methods that include degree of membership function method, k-nearest neighbor, decision tree classification, by using multi-scale segmentation method and the comparison of translation results and artificial visual solutions. The results of classification show that the three kinds of classification methods all can extract image information to a certain extent, but all of them has different characteristics: k-nearest neighbor classification method has the advantages of simple operation, need less samples than other method, the calculation speed is fast, but it get the lowest classification accuracy; the membership function classification method need to study the range of the image object, and can flexibly a membership function to express each classification, classification accuracy; decision tree classification need to fully understand the difference between each class, and can be used in a variety of conditions in the software language to express, and the decision tree should not be too deep, otherwise easily lead to the collapse of the decision tree, but due to its complex factors, the classification accuracy of the most trusted.