On The Applications Of Pattern Recognition Methods In Automatic Spectral Processing Of Celestial Objects

With the rapid development of astronomical observation, very large numbers ofspectra have been produced. After the completion of LAMOST project in China, about20,000 to 40,000 spectra will be collected at each observation night. With suchvoluminous data, automatic spectral analysis and recognition become desirable andnecessary for efficiency. Astronomical spectral analysis and recognition mainly includespectral classification, parameter measurement and spectral line identification. In thisthesis, we investigate the relative pattern recognition methods of the galaxy spectraprocessing. The main points of our work can be summarized as follows:1. For noise reduction of the astronomical spectra, an adaptive bilateral filteringmethod based on mean shift algorithm is presented. The filter is governed by thekernel width in spatial domain. Its kernel width in range domain is chosenadaptively by the local characteristic of the signal. It can remove impulsive noiseand improve smoothing of non-impulsive noise with edges preserved. The method isfit for the noise removal of non-emission-line spectra.2. Spectral line extraction of non-emission-line spectra is the most difficult task inspectral line auto-extraction of the astronomical spectra. Existing methods can dealwell with the emission-line spectra, but poorly with the non-emission-line spectra.We present a novel approach for non-emission-line spectra. It improves the qualityof continuum fitting by using wavelet transform iteratively. In addition, adaptivelocal thresholding associated with the universal thresholding is used to extractspectral lines. Experiment shows that this method is superior to the traditionalwavelet method.3. An algorithm for classification of galaxy spectra is proposed, which divides thegalaxy spectra into normal galaxy and active galaxy spectra. At first, a four-levelwavelet decomposition of the original spectrum is performed, and the waveletcoefficient at the fourth level, which mainly includes the information of spectrallines, is chosen as the wavelet feature of the spectrum. Secondly, principalcomponents analysis is used to compress the wavelet feature and to get therecognition feature of the spectrum. Finally, Fisher linear discriminant analysis isemployed for classification. Different from existing methods, this approach canrecognize the galaxy spectrum whose flux is uncalibrated and redshift is unknown.4. Redshift determination and spectral line identification are the most important partsin spectral recognition. An approach based on the knowledge of spectral lines ispresented. Firstly, definitions of the redshift candidate and spectral line candidateare given, and the candidates are cross-validated according to the definitions;Secondly, the density is estimated at every redshift candidate using the Parzenwindow technique;Finally, the average of redshift candidates in Parzen window ofthe redshift candidate with maximum density is the redshift, and the spectral linecandidates corresponding to those redshift candidates are spectral lines. Comparedwith other methods based on spectral line matching, this method has a lowerdependence on the quality of spectral line extraction.5. A novel technique based on cross-correlation for redshift determination is presented.Firstly, the spectral line candidates are extracted and used to determine the redshiftcandidates;then,the cross-correlation between the observed spectra and thetemplates shifted by every redshift candidate is measured, and the redshift candidatecorresponding to the largest correlation is chosen as the redshift of the observedspectra. The templates are constructed by PCA from the four templates of normalgalaxies and the seven templates of active galaxies offered by Kinney&Calzettirespectively. The advantage is that there is no need of the classification for theobserved galaxy spectrum. Compared with the traditional cross-correlation basedmethods, this method can be used for high redshift and high noise spectra.Compared with the methods based on spectral line matching, this method is lesssensitive to the quality of spectral line extraction.