Research On A New Algorithm Of Remote Sensing Image Intelligent Classification

With the rapid development of remote sensing technology, more and more remote sensing data are available for people, and the problem how to make full use of these large-scale data has caused wide public concern, and remote sensing image classification is one of the effective means to solve this problem. Now remote sensing image classification has been widely used in areas ranging from the detection of land use status, the dynamic prediction of land cover, to the establishment of remote sensing database, the thematic map making and such.In recent years, Support Vector Machine(SVM) algorithm demonstrates superior performance in remote sensing image classification, and it improves the theoretical weaknesses of the traditional Artificial Neural Network(ANN) algorithm by comparison, such as being limited to local minimum easily. Therefore, theoretical research on the SVM algorithm and its application in remote sensing image processing have become one of the hot research spots.In this paper, further research based on the existing remote sensing image intelligent classification algorithm is done. The main contents and innovation points are as follows:(1) In-depth study of the existing methods of remote sensing image classification, comparison between ANN and SVM classifier which are used most commonly in recent years shows that SVM classifier has superior performance. By making comparison of the characteristics and application ranges among the existing texture analysis methods of remote sensing images, Gray Level Co-occurrence Matrix (GLCM) with stronger generality and robustness has been chosen finally. In this paper, the inevitable problems that "the same objects with different spectrums" and "different objects with the same spectrums" in the remote sensing images have been better solved by extracting texture features from the TM image using the second order statistic of GLCM.(2) K-type SVM is applied in the TM multi-spectral remote sensing image classification for the first time. We can get a more superior classifier by changing the kernel function of SVM. Compared with the SVM based on radial basis kernel function, the generalization ability of K-type SVM is stronger, and its classification efficiency is higher.(3) This paper studies the combination of spectral features and texture features belonging to different principal components, and then put them as the input vectors of the SVM classifier. The simulation tests are done with TM remote sensing image, and the results show that when texture features of the first principal component and spectral features are imported into K-type SVM, a higher classification accuracy and a higher Kappa coefficient can be gotten, meanwhile, classification results achieve the best on the whole.(4) An active sample labelling algorithm based on the kernel fuzzy C-means clustering is proposed as the manual sample labelling method is relatively ineffective, thereby realizing the semi-supervised classification strategy improvement of SVM. Through the contrast simulation experiment of labelling samples actively based on fuzzy C-means, it is concluded that the active labelling method proposed in this paper is more effective.