Studies On Classification Method Of Celestial Spectral

LAMOST project is one of the National Major Scientific Projects during the period of the national ninth five-year plan. The novel design of LAMOST system is a telescope that provides a combination of a large aperture and a large field of view,both of which feed a highly-multiplexed spectroscopic system. LAMOST is a reflecting Schmidt telescope with its optical axis fixed along the north-south meridian.In fact, LAMOST is the telescope which has the highest spectrum-acquiring rate in the world currently, so we will collect as many as tens of thousands spectra per observation night by using LAMOST. In the face of the amazing spectral data from LAMOST, the traditional methods of processing spectra artificially are no longer applicable. Therefore, it needs us to study and find some methods which can be used to process the more spectra data automatically. Meanwhile, some classical data mining algorithms have been applied to classify celestial spectrum data increasingly.Based on above reasons, in this thesis we studied two methods which can be used by the automatically classification of spectral data. And, the main contributions of this thesis are as follows:(1) Pre-processing of spectral data and extraction of eigen-spectrum. The spectral data in this paper covering stars, galaxies and quasars are selected from the LAMOST-DR1 and SDSS-DR. This thesis adopts the PCA method to extract the spectral features after the pre-processing including de-noise and flux normalized working, that is, we take the subspace produced by eigen-spectrum as feature space which corresponds to eigenvalues which variance contributions are over 99 %.(2) Spectral classification based on KNN and WKNN, a advanced spectral classification method of dual distance-weighted K-nearest neighbor(DWKNN). The DWKNN is a modified algorithm based on KNN. In this thesis, this method is successfully is applied to the classification of stars, galaxies and quasars. The experimental results show that the classification accuracy among DWKNN towards the three different spectrums are 99.12%,98.73% and 93.57% respectively, and the accuracy is also superior to traditional KNN.(3) Spectral classification based on K-means-SVM. K-means method and support vector machine method are classical data mining algorithms. K-means method is easy to actualize without training which belongs to unsupervised classification, but it achieved not so good results in spectral classification. With respect to the SVM method, it needs training and has the better accuracy than K-means method through using manual training sample to establish classificationmodel, that will cause to much time and work. Therefore, the two methods have strong complementarities in theory. This paper will apply a semi-supervised classification method which combines K-means and SVM to data classification of galaxy and quasar in LAMOST-DR1 and star and galaxy in SDSS-DR8, and finally it obtains good results.