Support Vector Machines Ensemble And Its Application In Remote Sensing Classification

Support Vector Machines (SVM) has been applied in many fields and achieved plentiful fruits already since proposed by Vapnik in 1995. However, there are some defects with SVM. Firstly, approximate algorithms used to solve optimization problem degrade the generalization ability. Secondly, there are no good methods for the choice of optimal kernel function and parameters. Thirdly, multi-class SVM combining of several two-class SVMs also degrades the generalization ability. Ensemble learning is one of the hot topics of machine learning. And the ensemble based on neural network and decision tree has made great progress. However, it is quite new on the research of SVM ensemble. In this dissertation, the technology of SVM ensemble is studied and applied to the remote sensing classification. The main contribution of this dissertation are summarized as following:1. SVM is the relatively stable classifier in comparison with neural network and decision tree. Bagging SVM and Boosting SVM cannot improve the classification result. In this dissertation, RBaggSVM and RBoostSVM algorithms manipulating both training sets and SVM model parameters are proposed. It is characteristic of randomization of SVM model parameters within specific scope for achievement of diversity component classifiers.2. Seeking for part of classifiers for ensemble is regarded as the optimization problem. Genetic algorithm has the ability of global optimization. GABaggSVM and GABoostSVM algorithms of classifiers selection based on genetic algorithm are proposed. Genetic algorithm is used to optimize weights of components classifiers. Then selects the optimal part of component classifiers for ensemble.3. DERBaggSVM and DERBoostSVM algorithms of dynamic ensemble based on local accuracy are proposed. Parts of component classifiers are selected...