Research On Remote Sensing Image Classification Methods Based On Kernel Theory

These problems of nonlinearity, computational complexity, fuzziness and few labeled data exist in remote sensing image classification. In this thesis, several algorithms, such as generalized discriminant analysis (GDA), fuzzy C-means (FCM), and spectral angle match (SAM), are extended to their kernel patterns by introducing the kernel method, the semi-supervised learning technique, and the neighbor sample characteristic, et al. Kernel-based framework for remote sensing image classification is constructed. Those new kernel pattern related algorithms are applied in training data reduction, nonlinear feature extraction, and remote sensing image classification for improving the classification accuracy and efficiency, and reducing the computational complexity. The main work and results are as follows:1. A method is proposed to reduce training data of remote sensing image classification in large datasets with support vectors from nonlinear support vector machine (NSVM). The NSVM method reduces training data and computational complexity of training classifier while retaining the generalization of the classifier.2. Nonlinear feature extraction has high computational complexity in large datasets. A greedy GDA (GGDA) method is proposed to reduce training data and deal with the nonlinear feature extraction problem, and used in data of remote sensing image. The simulation results indicate that the feature extraction performance of both GGDA and GDA methods outperforms one of these compared methods. In addition of retaining the performance of the GDA method, the GGDA method reduces the computational complexity of the nonlinear feature extraction in large datasets.3. These problems of nonlinearity, fuzziness and few labeled data are rarely considered in traditional remote sensing image classification. A semi-supervised kernel fuzzy C-means (SSKFCM) algorithm is proposed to overcome these disadvantages of remote sensing image classification. On the one hand, with the kernel method, the input data is mapped implicitly into a high-dimensional feature space where the nonlinear pattern now appears linear in the SSKFCM algorithm. On the other hand, by the semi-supervised learning technique, the SSKFCM algorithm combines the labeled and unlabeled data together to improve the classification accuracy of remote sensing images.4. Both FCM and kernel FCM (KFCM) algorithms have a disadvantage of equal partition trend for data sets with minimizing the sum of error squares objective function. Several weighted FCM (WFCM) and weighted KFCM (WKFCM) algorithms are proposed to overcome the disadvantage of FCM and KFCM, by involving the neighbor sample density (NSD), the neighbor sample membership (NSM), and both the neighbor sample density and membership (NSDM) into the FCM and KFCM algorithms, respectively. The weighted coefficient of the NSD exerts an influence on the sum of error squares objective function, the higher the value, the larger the influence; on the other hand, the neighbor samples have the tendency of the approximately same membership value by the weighted coefficient of the NSM. Experimental results indicate that these weighted algorithms improve the classification performance to some extent, and significantly improve the unsupervised classification capacity of remote sensing images compared with FCM and KFCM.5. A kernel spectral angle match (KSAM) algorithm is proposed to deal better with the nonlinear classification problem of remote sensing image. The KSAM algorithm extends the spectral angle match (SAM) algorithm by introducing the kernel method. Experimental results indicate that the classification accuracy of the KSAM algorithm is superior to one of the SAM algorithm. Experimental results indicate still that kernel parameters of poly and sigmoid kernel are excessively sensitive, and a narrow bound of kernel parameters can be chosen for the optimal classification; the classification performance of ERBF and RBF kernel is superior to one of Poly and Sigmoid kernel, and a wide bound of kernel parameters in ERBF and RBF kernel can be chosen for the optimal classification in the KSAM algorithm.6. Comprehensive classification experiment is accomplished to validate further the classification performance of these proposed kernel pattern classification algorithms. The experiment results indicate that the classification performance of SSKFCM, NSDM-WKFCM and KSAM is superior to one of the same type compared algorithm.