LAMOST (Large Sky Area Multi-Object Fiber Spectroscopic Telescope) project is one of a few key scientific projects during the period of the national ninth five-year plan. After its expected completion at the end of 2005, about 10,000~20,0000 spectra will be collected per observation night. LAMOST project urgently needs a fully automatic spectral processing and analysis system due to its voluminous data, a dataset of up to 107 spectra. To this end, this work is particularly focused on finding and designing suitable techniques for spectral classification and radial velocity redshift measurements. The main contributions are following:(1) Spectra classification techniques based on local PCA (LPCA) and Kernel PCA (KPCA). Two classification algorithms based on LPCA and KPCA are proposed for the classification of stars, galaxies and quasars. Experiments show that LPCA is capable of extracting more pieces of useful information on stars and quasars than the original PCA does, and as a result, the corresponding correct classification rate is higher. It is particularly suitable for large-scale spectra processing thanks to its high computational efficiency. Experiments show that KPCA reaches its best performance when the width of Gauss window equals 2. Comparatively, KPCA performs slightly better than PCA does for classification, and both reach their best result when the number of he principal components is fixed to 20. These experimental conclusions are useful for the design of future LAMOST. (2) Spectra classification based on kernel trick. (a) A kernel based Generalized Discriminant Analysis (GDA) technique is proposed for the classification of stars, galaxies, and quasars. LDA, GDA, PCA and KPCA are experimentally compared with these 3 different kinds of spectra. Among these 4 techniques, GDA obtains the best result, followed by LDA, and PCA is the worst one. Although KPCA is also a kernel based technique, its performance is not satisfactory if the selected number of the principalcomponents is small, and in some cases, it appears worse than LDA, anon-kernel based technique. (b) A kernel based covering algorithm, called the kernel covering algorithmis proposed. This algorithm is a combination of kernel trick with the coveringalgorithm, and is used to extract the support vectors in feature space. Theexperiments show that although the kernel covering algorithm has a comparableclassification performance compared with the covering algorithm, the numberof its resulting support vectors is significantly smaller than that of the coveringalgorithm.
修改评论