Pattern‐Recognition Methods in Automatic Techniques for Astronomical Spectral Analysis

Abstract

The Large Sky Area Multi-Object Fiber Spectroscopic Telescope (LAMOST) is developing rapidly in China, and it will yield a very large number of fiber spectra in the near future. To cope with the anticipated flood of spectrographic data, automated methods for reducing the spectra, measuring spectral features, and classifying them are being investigated. First, this dissertation describes the present status of automated methods for analyzing astronomical spectra in other spectral surveys, including principal component analysis (PCA), artificial neural network (ANN), wavelets, and other methods. For example, B. C. Bromley et al. (1998, ApJ, 505, 25) applied PCA for spectra from the Las Campanas Redshift Survey, S. R. Folkes et al. (1999, MNRAS, 308, 459) and S. R. Folkes, O. Lahav, & S. J. Maddox (1998, ApJ, 492, 98) classified spectra from the 2dF Galaxy Redshift Survey using PCA and ANN, and F. J. Castander et al. (2001, AJ, 121, 2331) applied PCA and wavelets in Coma data obtained by the Sloan Digital Sky Survey. One of the most difficult problems for LAMOST is that automated spectral processing must carry out both classification and redshift determinations, so simple independent pipelines will not work. For this thesis, the author has built a pattern-recognition–based automated frame for processing spectral data. According to the widely used pattern-recognition process, the work involves three aspects: preprocessing, feature extraction, and classification determination. The preprocessing step is to fit the spectral continuum. Feature extraction is the key step in the pattern-recognition system, including measuring and identifying spectral features such as lines and the 4000 A break and computing redshifts. Finally, automated classification should make use of the information from spectral lines. A.-L. Luo & Y.-H. Zhao (2000, Acta Astrophys. Sinica, 20, 427; 2001, Chinese J. Astron. Astrophys., 1, 563) described a continuum-fitting method that has been used in this dissertation. Instead of using a fixed wavelet scale (J.-L. Starck, R. Siebenmorgen, & R. Gredel 1997, ApJ, 482, 1011), the author uses a wavelet filter bank to subtract the continuum in a spectrum. A. Luo & Y. Zhao (2001, Spectrosc. Spectral Anal., 21, 19) suggested steps for measuring spectral features and computing redshifts. A multiresolution feature-extraction method is introduced in this dissertation. Generally, automated classification needs blueshifted spectra, classifying them in a space such as PCA space. K. Glazebrook, A. R. Offer, & K. Deeley (1998, ApJ, 492, 98) developed an automated redshift determination method called PCAZ. The author has corrected some errors inherent in this method and has presented an improved approach. The author has also applied statistical learning theory based on the Support Vector Machine to accomplish the classification and search for strange objects. When the signal-to-noise ratio (S/N) is too low, noise will cause decreasing accuracy in the classification. The noise is not Gaussian and varies with wavelength. The results of continuum fitting, feature extraction, and classification depend heavily on estimating the noise. In this dissertation, the author has applied the wavelet-based hidden Marcov model (HMM) to estimate the noise distribution in low S/N spectra (A. Luo & Y. Zhao 2001, Nuovo Cimento B, 116, 879) and has improved M. S. Crouse et al.’s (1998, IEEE Trans. Signal Process., 46, 886) tying method to increase the training data to make the HMM more robust and to estimate the noise more precisely.