Research On Multi-spectral Remote Sensing Image Classification Based On The Hybrid Model Of The SVM And K-means

With the rapid development of remote sensing data acquisition technology, the diversity of data source is far ahead of the processing level of remote sensing data.Therefore, the research on remote sensing data processing and analysis methods has vital significance on promoting the application of remote sensing data. The classification of multispectral remote sensing data is one of the most challenging directions in the field. As one of the most effective way of statistical learning theory, The Support Vector Machine, which can overcome the problem of over-fitting, non-liner, curse of dimensionality and local minimum comparing with the traditional methods, is becoming a research hotspot among numerous remote sensing image classification methods.SVM is a machine learning algorithm for multi-spectral remote sensing image classification, nearly two decades of history. As a kind of classical clustering algorithm, k-means algorithm has also been used to classify all kinds of data. We can find that both of the two algorithms has their own unique advantages and disadvantages by reading a large number of articles.Therefore, in order to meet the applications requirements of the results image of multispectral remote sensing image classification, the classification performance of these two algorithms have some room for improvement.The training samples need to be selected to create a classification model before the classification, because SVM algorithm is a supervised classification method. However, k-means clustering algorithm is an unsupervised classification method, which can automatically make the unknown samples close to their clustering center without training samples. Consequently, these two classification algorithms have highly complementary in principle. In order to further improve the classification accuracy, this paper presents a hybrid classification model of SVM and k-means for multi-spectral remote sensing image classification. The mathematical derivation of the model has been finished and this strategy has been applied to the classification of Landsat TM multi-spectral remote sensing image data successfully. The working principle of the hybrid model is that:first using k-means algorithm to make the unknown samples close to their clustering center, and then choosing the training sample from the respective clustering center around a certain radius automatically, as training data of SVM.The experiments show that the classification accuracy of SVM model created by human marking samples is89.47%, while the classification accuracy of the hybrid classification model is97.53%. Obviously, the classification accuracy of the hybrid model has been improved significantly, so the effectiveness of the proposed method has been illustrated.In the process of research, a series of theoretical and practical problems included in the "hybrid model" has been analyzed in detail. Such as:the adaptation problem of the high dimensional data and the low dimensional data with different kernel function, The selection of SVM parameter optimization algorithm, the classification performance of SVM deformation algorithm, the solution of the quadratic programming problem and so on.In addition, this article also focuses on the influence caused by parameters of RBF kernel function and penalty factor on the performance of SVM classifier.Around the goal of this subject, this article has carried on five aspects of the research and a series of theoretical issues included in this article have been verified by the16experiments. Some important experiment data and conclusions have been obtained through the experiment, which enriched the content of theoretical research greatly. The research contents provide strong support for further theoretical research and laid the foundation for the research of the hybrid model multispectral remote sensing image classification model of SVM and k-means proposed in this paper.