Research On High Resolution Remote Sensing Image Building Extraction Based On The Multi-kernel Support Vector Machine

The rapid development of remote sensing technology improved the resolution of remote sensing image and increased data volumes dramatically, so more rich feature information contained in the image can be extracted and analysed more accurately. However, the accuracy of automatic extraction of high resolution remote sensing image information is not high presently, the building as a kind of extremely important artificial object target, extracting its all kinds of information better is of great significance for the promotion of high resolution remote sensing image in the identification and classification of target, the urban land planning and management and the like.Classification is a key problem in the extraction of image information, support vector machine(SVM) classification based on the kernel function is an effective method to solve the problem of image target classification. For the problem of the classification performance based on single-kernel SVM model have big differences and cannot extract the building accurately due to the different choice of kernel function and setting of kernel parameter, this paper presents the multi-kernel learning classification model of support vector machine(MKSVM) to extract the image of the building. This MKSVM classification model is a classifier which aimed at different features has different contribution to the extraction of different buildings, built by different function kernel be linear additivity through the way of weight, which the model contains all the basic characteristics of the kernel function, has excellent learning ability, generalization performance and flexibility.This research underneath Visual studio 2013 platform based on C++ language of ORFEO TOOLBOX(OTB) open source library distributed image processing algorithms, developed high resolution remote sensing image building information extraction system on the basis of the object-oriented, achieved using two segmentation and classification method to extract buildings. Firstly, utilized the secondary region merge of improved watershed method to segment images, meanwhile, use K-nearest neighbor supervised classification based on spectral feature extracted impervious surface merely included buildings, roads and cement square and the like, extract only blue workshop among the image in this process simultaneously; Then apply mean shift segmentation to impervious surface, according to the extracted spectral, texture and spatial features of building, input sample into MKSVM classification mathematical model that constructing by polynomial kernel function and radial basis function training to obtained classifier; finally, extract building through the training of MKSVM classifier. Compared the accuracy with the results of single-kernel SVM classification extraction buildings, relative to RBF kernel, two regions of MKSVM users classification precision is increased by about 1.5%, relatively to POLY kernel, two regions of MKSVM users classification precision is increased by about 3% to 5%; the Kappa coefficient of two regions is increased to 0.8579 and 0.8415, respectively.The experimental results show that, the operation of the image building information extraction system in this paper is reliable, object-oriented classification method based on MKSVM possesses higher accuracy when extracting buildings and better classification ability compared with single-kernel SVM extraction buildings.