Research On LSTSVM-based For Hyperspectral Image Classification

Object classification technology is an important branch of hyperspectral remote sensing image processing, and with the rapid development of society, the requirement of classification accuracy and efficiency is gradually increasing. The goal of the hyperspectral image classification is to distinguish the land-covers contained in the image, and assign the classes to each pixel according to their characteristics, nowadays,classification method is becoming more and more rich, however, the hyperspectral image classification technology is facing key issues: (1) The dimension of hyperspectral image data is general high,which leads to large computation; (2) Some classification algorithms, including the standard SVM, select a certain number of samples from the interested object category randomly to train the classifier and they don't consider the difference between each training sample, which leads to the redundancy in the training sample set and the low classificaiton efficiency; (3) The existing classification model which fusion the spatial and spectral information usually only use one kind of spatial information and they don't make full use of the spatial information in the original hyperspectral image, such as fusing the spectral information and texture information or structure information, and so on. Considering the above problems, this paper mainly studies the hyperspectral image classification algorithm based on Least Squares Twin Support Vector Machine(LSTSVM), and the "-1" class training samples which lay the center area are redeced by calculating on the basis of one-against-rest (1-a-r) multi-classifier structure, which improvs the classification efficiency; Meanwhile, in order to make full use of the information in the original hyperspectral image, the texture information ?structural information and spectral information are fused together for LSTSVM classification. The experiments show that the modified algorithm can not only improve the classification efficiency but also improve the classification accuracy. The main contributions are listed as follows:Firstly, the background and significance of hyperspectral image research are introduced.And the development of classification technology at home and abroad is also introduced;Secondly, the LSTSVM classification algorithm is mainly introduced on the basis of the introduction of standard Support Vector Machine (SVM) and Twin Support Vector Machine(TWSVM ). The experimental results show that the LSTSVM algorithm performances better than standard SVM;Thirdly, the Sample Reduction LSTSVM improved classification algorithm(SR-LSTSVM) on the basis of one-against-rest (1-a-r) multi-classifier structure is proposed.Firstly, the distance between each "-1" class training sample and its corresponding center is calculated, and they are sorted in ascending order according to the distance value. Then the"-1" class traing samples which have lower distance value are removed basing on the pre-reduction rate. Finally use the redeced "-1" class traing sample set and the original "+1"class traing sample set to train LSTSVM classifier and classify. Simulation results show that the improved classification algorithm can greatly shorten the classification time and without affecting the classification accuracy basically;Fourthly, an improved LSTSVM classification model basing on the fusion of spatial and spectral information is proposed. Firstly, the Principal Component Analysis (PCA) algorithm is used to implement principal component transformation to the original data, and the first three principal components are stored. Then the Extended Morphological Profiles (EMP) data is obtained by using four different sizes of circular Structure Elements (SE) to undertake opening operation and closing operation on the first three principal components, and the EMP data includes structural information and most of the spectrum information of original hyperspectral image. And the texture information is extracted from the first principal component by using Gabor filter. At last, fuse the EMP data and Gabor texture data for LSTSVM classify. Simulation results show that the improved classification model can not only improve the classification accuracy, but also lower the data's dimension, which reduces the computational complexity to some extent;Finally, the sample reduction strategy and improved LSTSVM classification model are combined together to form a new improved sample reduction LSTSVM classification algorithm. Experimental results show that the improved algorithm can not only improve the classification efficiency,but also improve the classification accuracy.