The Classification Of High Resolution Remote Sensing Images Based On Local Support Vector Machine

With the rapid development of wireless communication technology, the remote sensing image presents the trend of high resolution. The improvement of high resolution can provide more valuable information, and promote the development of various applications. On the other hand, the processing of the remote sensing images with high resolution have brought more and more challenges. With the continuous improvement of the resolution on remote sensing images, the demand for the accuracy of the application is also increasing, which leads to an exponential increase in the amount of data needed to be processed. Therefore, the image processing mode of the traditional artificial visual interpretation has been completely unable to meet the real needs. And the use of data mining, machine learning and so on which are used to classify images with computer automatically have become the main approaches in the field of remote sensing image processing. However, it is also time consuming and lower accuracy to process high resolution remote sensing images with computer automatically due to the explosive growth of data. At the same time, the high resolution remote sensing images has more features than traditional ones because it includes the spectral information which is the same as the general features and also have thedifferent features as texture features, geometric features and context information and so on. The traditional remote sensing image processing is mainly based on the gray scale value of the color classification, but it is lower accuracy of classification for the sake that there are phenomenon of "spectrum" and "synonyms spectrum" when the spectral feature is the only category for classification.Support vector machine (SVM) model is constructed by the support vectors in the training dataset. Therefore, it has the advantages of sparse, higher accuracy and faster running speed in classification, which is especially suitable for classifying the problems with small training data, nonlinear and high dimensional features. Based on its features, it is also very appropriate to classify the high resolution remote sensing images with mass data. However, it is not global consistency which may lead to lower accuracy. In order to improve performance of SVM, the algorithm called as Local Support Vector Machine combines KNN and SVM is applied in application. Due to the combination of the two classifiers, the LSVM has higher accuracy than these two.In order to further improve the accuracy of local support vector machine classification and reduce the complexity of its computation based on the features of high resolution remote sensing images, the paper makes a deep research on this problem based on existing documents.1. The improved KNNSVM local support vector machine algorithm based on uncertainty (BKNNSVM) is proposed. The problem of the time complexity of the algorithm KNNSVMhas been analyzed in detail. It is discovered that the unlabeled samples are determined by the two classifiers KNN and the SVMafter a thorough study of the KNNSVM classification.According to the relative experimental results of the time consuming of KNNSVM algorithm, it is found that the more unlabeled samples are determined by the SVM classifier, the more time is needed. Due to each unlabeled sample has a neighbors set which is met to binomial distribution, the uncertainty of each unlabeled sample can be estimated by the Beta distribution which is the conjugate prior distribution of the binomial distribution. As a result, the BKNNSVM algorithm is proposed based on the definition of Beta distribution, the distribution curves of different parameters, and the characteristics of KNNSVM. The algorithm has two parameters of uncertainty threshold and the number of neighbors K which can control the increase (decrease) number of unlabeled samples classified by KNN to decrease (increase) the time consumption of BKNNSVM. The experimental results show that the algorithm BKNNSVM can keep the accuracy of KNNSVM with lower time complexity.2. A local support vector machine algorithmbased on distance (DLSVM) is proposed. With the analysis of experimental results from wrong classified samples, it is discovered that samples which are near to the hyperplane have more errors and lower accuracy than others and the accuracy of them will be improved when the classifier is the local support vector machine. In order to identify these samples easy to be wrong classified, we borrow the idea of active learning method. The active learning approach is to improve the accuracy of the classifier by searching for the unlabeled samples with most information, and among them the samples selected by uncertainty is the same with us. Combing the active learning method based on distance and the analysis of the wrong classified samples of SVM, we proposed the local support vector machine algorithm based on distance DLSVM. With this algorithm, the distance of the unlabeled samples and the hyperplane have been calculated and the samples whose distance is less than the threshold are used to build the local support vector machine. With the experiments on high resolution remote sensing images, due to the local support vector machines are built only by samples whose distance is less than threshold, the time consumption improved less and the accuracy of classification improved obviously. Especially, the time complexity of the DLSVM is far less than KNNSVM. 3.The algorithm of directed acyclic graph local support vector machine (DAGLSVM) is proposed for multiple classification. Support vector machine classifier SVM is applied for binary classification. It is extended for multiple classification for the sake that it has high accuracy and powerful generalization ability when it is used for binary classification. The main approaches of SVM for multiple classification include one VS one, one VS many, the Directed Acyclic Graph Support Vector Machine (DAGSVM) and SVM based decision tree and so on. Considering that it will take much time when the algorithm KNNSVM is applied for multiple classification directly because the number of the local SVM needed will square more than for binary classification. Therefore, in order to decrease time complexity, the algorithm of the directed acyclic graph local support vector machine (DAGLSVM) which combines the features of the highresolution remotesensing images and the algorithm DAGSVM is proposed. The algorithm builds a nearest neighbor support vector training sample set for each unlabeled sample, and the maximum and minimum of two classes are selected to establish a support vector machine, which can avoid the errors accumulated from the top due to mistaken selecting the two classes for SVM. The training samples whose label are different from the one classified by the SVM will be eliminated from the training dataset. And the iteration will be repeated until there is only one class remaining in the training dataset. As a result, this class is the unlabeled sample belonging to. Experimental results show that it can effectively improve classification accuracy of the highresolution remotesensing images when the K is suitable with little incremental running time.4. The image processing platform has been designed which is applied to finish all the experiments in the paper. Based on the open source code LIBSVM, which is developed by the Professor Lin Zhiren of National Taiwan University, the classification platform with software of remote sensing image support vector machine is designed by JAVA. The software combining with WEKA and LIBSVM realizes the classified algorithms of KNN, KNNSVM, BKNNSVM, DLSVM, DAGSVM and DAGLSVM. The abilities of display of remote sensing image data, image feature extraction, classification, classification results storage and the classified results reveal are realized in the platform.In conclusion, due to the local support vector machine can not only improve the accuracy of traditional support vector machine, but also maintain its good generalization and support small training dataset, a variety of optimization algorithms based on the local support SVM which can improve the classification performance of the high resolution remote sensing images, reduce the time complexity of computing are proposed in the paper. These algorithms is helpful to further improve the processing speed of remote sensing image, and play a greater social and economic...