[The Automatic Recognition and Detection of Sky-Subtraction Residual Componentin the Stellar Spectra].

The skylines, superimposing on the target spectrum as a main noise, will reduce the signal-to-noise ratio of the spectrum. If the spectrum still contains a large number of high strength skylight residuals after sky-subtraction processing, it will not be conducive to the follow-up analysis of the target spectrum. At present, the study on the automatic recognition of the abnormal sky-subtraction stellar spectra is limited in number. We can only find the abnormal sky-subtraction spectra by manual inspection, and this will reduce the speed of detection. This paper analyzes the influence factors of sky-subtraction results and finds the characteristics of the abnormal sky-subtraction spectra. A simple and effective method is proposed to automatic recognize the abnormal sky-subtraction stellar spectra which have been processed with the LAMOST Pipeline processing procedure and find the positions of the abnormal skylines. In this method, all the spectra are normalized first; the abnormal skyline is determined by detecting whether there exits any high strength skyline residuals which are similar to the emission line or absorption line. Finally, all the abnormal skyline positions in the spectra are obtained in this method. The experimental results with the LAMOST spectroscopic dataset show that this method can recognize the abnormal sky-subtraction spectra and find the abnormal skyline positions of different residual strength effectively. In addition, the method is simple and has high recognition efficiency, and can be applied to the automatic detection of abnormal sky-subtraction of large number of spectra.

[Stellar Alpha Element Abundance Estimation Using LASSO Algorithm].

In this paper, a new method based on LASSO algorithm is studied for the estimation of stellar alpha element abundance. The information of alpha elements (O, Mg, Ca, Si, and Ti) of massive stars will help us to better understand the evolution of the galaxy. Presently the main method of determining the alpha element abundances from the low resolution spectra is the template matching method. However, it is difficult for us to optimize the algorithm parameters and the algorithm is sensitive to the noise. Thus, it is necessary to study the new method to determine the abundance. The experimental results show that the accuracy of LASSO algorithm on ELODIE spectra is 0.003 (0.078) dex. To explore the impact of the spectral resolution variation, we use ELODIE spectra to generate the spectral data sets with following resolutions: 42 000, 21 000, 10 500, 4 200 and 2 100 by using the Gaussian convolution. The results of the LASSO algorithm on these data sets are 0.003 3 (0.078) dex, -0.05 (0.059) dex, -0.007 (0.069) dex and -0.004 5 (0.067) dex, respectively. These results show that the LASSO algorithm is not sensitive to the change of the resolution. In order to verify the robustness of LASSO algorithm against the change of SNRs, we use ELODIE to generate the spectral data sets with following SNRs: 30, 25, 20, 15 and 5. The results of LASSO algorithm on the above data sets are: -0.002 (0.076) dex, -0.090 (0.073) dex, 0.003 6 (0.075) dex, 0.007 6 (0.078) dex and -0.009 (0.080) dex, respectively. Thus, LASSO algorithm is not sensitive to the change of SNR. Therefore, the LASSO algorithm is suitable for low resolution and low SNR spectra such as LAMOST and SDSS spectra. The accuracy of Lasso algorithm on the SDSS spectra is 0.003 7 (0.097) dex, and the results of LASSO on globular and open clusters show good agreement with literature values (within 1σ). Therefore, the LASSO algorithm can be used to estimate the alpha element abundances of stars.

Outlier Data Mining and Analysis of LAMOST Stellar Spectra in Line Index Feature Space.

Large scale spectrum survey will produce mass spectral data and offer chances for searching rare and unknown types of spectra, which is contribute to revealing the evolution law of the universe and the origin of life. Data mining in outlier data in sky survey can serve the purpose of finding special spectra. Line index can be used in spectra data dimension reduction, keeping the spectral physical characteristics as much as possible, and at the same time, it can effectively solve the high dimensional spectral data clustering analysis in the high computation complexity. This paper proposed a method outlier data mining and analysis for massive stellar spectrum survey data based on line index characteristics, according to this, an outlier spectral data analysis method was proposed using line index characteristics space. Experimental results demonstrated that (1) using line index as the characteristic value of the spectrum can quickly perform the outlier data mining for high dimensional spectral data, and it can solve the problem of high computation complexity of the high dimensional spectral data. (2) this outlier data mining method was conducted based on the clustering results; it can effectively finding out emission stars, late type stars, late M type stars, extremely poor metal stars, and even finding spectra data missing certain data. (3) outlier data mining in line index feature space can help to analysis of rules of special stars found in the feature space. The mothed proposed in this paper based on the characteristics of line index outlier data mining and analysis method can be applied to the study of survey data.

[Research on M Dwarf Sub-Classification Based on the Measurement of Residual Distribution].

The classification of stellar spectra is an important job in data processing of astronomy, which is mainly used for searching celestial spectra with known types in massive data survey. This paper focuses on LAMOST M dwarfs fine classification based on measurement of residual distribution. Residual distribution measurement is a measurement method used to measure the distance between two spectra. In the process of calculating the distance between two spectra, normalized processing should come first. Then the residuals of the sampling points of corresponding wavelength are calculated. Eventually the standard deviation of the residual distribution as the distance between the spectra is calculated. In this paper, the M star of LAMOST DR2 is used as the experimental data of classification. The experimental results show that the spectra data can be classified more accurately with the measurement method of residual distribution than the use of other traditional classification methods. The effect of spectral classification is affected by signal to noise ratio, outliers, residual standardized coefficient and other factors.

[Research on the Clustering of Massive Stellar Spectra Based on Line Index].

Clustering algorithm is an important algorithm used to find the data distribution and implicit scheme in data mining. It can study spectra of large amount, multi-parameter and categories unknown simply and effectively. Using lick index as the eigenvalues of spectra can effectively improve the speed to calculate the high-dimensional spectra which can also retain more astrophysical characteristics of spectra. This paper finishes clustering of the survey data with k-means algorithm, using lick index as the eigenvalues of data with finished analysis results. The results show that the new method can gather data with similar physical characteristics together quicker and efficiently, with very good results in discovering rare stars. This method can be applied to the study of Survey data.

[A New Redshift Measuring Method for Low-Quality Galaxy Spectra Based on Multi Resolution Fusion Distance].

The redshift measurement of galaxy spectrum is a key issue in large astronomical spectral survey. Its goal is to extract the redshift from spectrum, which is caused the Doppler Effect. With the development of the extragalactic sky survey project, the distance (redshift) of the observed targets is becoming further. As a result, the magnitude of the observed objects becomes darker and the spectral quality becomes poorer. Therefore, how to effectively and accurately measure the redshift from these low quality spectra is becoming an important problem in the extragalactic survey. Considering the spectral features and the data character, a new definition of multi-resolution fusion distance for low quality spectra is proposed. In this paper, we put forward a redshift measuring method for low quality galaxy spectra. This method combines the spectral features with different resolutions. The template spectrum and the spectrum to measure are reduced to the resolution and then a distance is computed by combining the offset of the above two spectra in different wavelengths. Then, a fusion distance is weighted averaged from the distances with different resolutions. In this paper, the effect of signal-to-noise ratio (SNR) on the measuring accuracy of the proposed method is discussed. The measuring accuracy is larger than 90% when the SNR is larger than 5. A large number of experiments show that the method proposed in this paper is very efficient in measuring the redshift of the low-quality galaxy spectra and the measuring error has nothing to do with the redshift value. The proposed method can be applied in redshift measurement of galaxies for the large-scale survey data.

[A Method of Stellar Spectral Classification Based on Map/Reduce Distributed Computing].

Celestial spectrum contains a great deal of astrophysical information. Through the analysis of spectra, people can get the physical information of celestial bodies, as well as their chemical composition and atmospheric parameters. With the implementation of LAMOST, SDSS telescopes and other large-scale surveys, massive spectral data will be produced, especially along with the formal operation of LAMOST, 2 000 to 4 000 spectral data will be generated each observation night. It requires more efficient processing technology to cope with such massive spectra. Automatic classification of stellar spectra is a basic content of spectral processing. The main purpose of this paper is to research the automatic classification of massive stellar spectra. The Lick index is a set of standard indices defined in astronomical spectra to describe the spectral intensity of spectral lines, which represent the physical characteristics of spectra. Lick index is a relatively wide spectral characteristics, each line index is named after the most prominent absorption line. In this paper, the Bayesian method is used to classify stellar spectra based on Lick line index, which divides stellar spectra to three subtypes: F, G, K. First of all, Lick line index of spectra is calculated as the characteristic vector of spectra, and then Bayesian method is used to classify these spectra. For massive spectra, the computation of Lick indices and the spectral classification using Bayesian decision method are implemented on Hadoop. With use of the high throughput and good fault tolerance of HDFS, combined with the advantages of MapReduce parallel programming model, the efficiency of analysis and processing for massive spectral data have been improved significantly. The main innovative contributions of this thesis are as follows. (1) Using Lick indices as the characteristic to classify stellar spectra based on Bayesian decision method. (2) Implementing parallel computation of Lick indices and parallel classification of stellar spectra using Bayesian based on Hadoop MapReduce distributed computing framework.

[A Method to Estimate Metal Abundance from Stellar Spectra Using Ca Line Index].

This paper presents a method to estimate stellar metallicity based on BP neural network and Ca line index. This method trains a BP ANN model from SDSS/SEGUE stellar spectra and parameters provided by SSPP. The values of Teff and the line index of Ca lines are the input of network while the [Fe/H] values are the oputput of the network. A set of samples are resampled from the set of all and then a network model is trained. The network can be used to predict the stellar metallicity from low-resolution spsectra. The experiment shows that the proposed method can accurately and effectively measure the [Fe/H] from the stellar spectra.

[A New Distance Metric between Different Stellar Spectra: the Residual Distribution Distance].

Distance metric is an important issue for the spectroscopic survey data processing, which defines a calculation method of the distance between two different spectra. Based on this, the classification, clustering, parameter measurement and outlier data mining of spectral data can be carried out. Therefore, the distance measurement method has some effect on the performance of the classification, clustering, parameter measurement and outlier data mining. With the development of large-scale stellar spectral sky surveys, how to define more efficient distance metric on stellar spectra has become a very important issue in the spectral data processing. Based on this problem and fully considering of the characteristics and data features of the stellar spectra, a new distance measurement method of stellar spectra named Residual Distribution Distance is proposed. While using this method to measure the distance, the two spectra are firstly scaled and then the standard deviation of the residual is used the distance. Different from the traditional distance metric calculation methods of stellar spectra, when used to calculate the distance between stellar spectra, this method normalize the two spectra to the same scale, and then calculate the residual corresponding to the same wavelength, and the standard error of the residual spectrum is used as the distance measure. The distance measurement method can be used for stellar classification, clustering and stellar atmospheric physical parameters measurement and so on. This paper takes stellar subcategory classification as an example to test the distance measure method. The results show that the distance defined by the proposed method is more effective to describe the gap between different types of spectra in the classification than other methods, which can be well applied in other related applications. At the same time, this paper also studies the effect of the signal to noise ratio (SNR) on the performance of the proposed method. The result show that the distance is affected by the SNR. The smaller the signal-to-noise ratio is, the greater impact is on the distance; While SNR is larger than 10, the signal-to-noise ratio has little effect on the performance for the classification.

[SKLOF: a new algorithm to reduce the range of supernova candidates].

Supernova (SN) is called the "standard candles" in the cosmology, the probability of outbreak in the galaxy is very low and is a kind of special, rare astronomical objects. Only in a large number of galaxies, we have a chance to find the supernova. The supernova which is in the midst of explosion will illuminate the entire galaxy, so the spectra of galaxies we obtained have obvious features of supernova. But the number of supernova have been found is very small relative to the large number of astronomical objects. The time computation that search the supernova be the key to weather the follow-up observations, therefore it needs to look for an efficient method. The time complexity of the density-based outlier detecting algorithm (LOF) is not ideal, which effects its application in large datasets. Through the improvement of LOF algorithm, a new algorithm that reduces the searching range of supernova candidates in a flood of spectra of galaxies is introduced and named SKLOF. Firstly, the spectra datasets are pruned and we can get rid of most objects are impossible to be the outliers. Secondly, we use the improved LOF algorithm to calculate the local outlier factors (LOF) of the spectra datasets remained and all LOFs are arranged in descending order. Finally, we can get the smaller searching range of the supernova candidates for the subsequent identification. The experimental results show that the algorithm is very effective, not only improved in accuracy, but also reduce the operation time compared with LOF algorithm with the guarantee of the accuracy of detection.

[Stellar spectral outliers detection based on Isomap].

How to find the spectra misclassified by traditional methods is the key problem that has been widely studied by the experts of astronomical data processing. We found that Isomap algorithm performs well for this problem. By comparing the performance of Isomap with that of principal component analysis (PCA), we found that (1) Isomap can project the spectra with similar features together and project the spectra with different features far away, while PCA may project the spectra with different features into nearby regions; (2) the outliers given by Isomap can be easily determined, and most of the outliers are binary stars with high scientific values; while the outliers given by PCA are difficult to determine and most of outliers are not binary stars. Thus, Isomap is more efficient than PCA in finding the outliers. Since the spectral data used in experiment are the spectra from the ninth data release of Sloan Digital Sky Survey (SDSS DR9), Isomap can find the spectra misclassified by SDSS pipeline efficiently and improve the classification accuracy obviously. Furthermore, since most of the spectra misclassified by SDSS pipeline are binary stars, Isomap can improve the efficiency of finding the binary stars with high scientific values. Though the experiment results show that Isomap is more sensitive to the noise than PCA, this disadvantage will not affect the application of Isomap in spectral classification since most of the spectra with low signal-to-noise ratios are the spectra whose spectral type can't be determined manually.

[Application of KPLSR based on line index in stellar atmospheric physical parameter measurement].

In the present paper, the kernel partial least squares regression (KPLSR) method was used to measure the atmospheric physical parameters (effective temperature, surface gravity, an d chemical abundance) based on the use of Lick line index. The proposed method can reduce the computation cost and achieve an ideal measure precision. At first, the Lick indices of Kurucz synthetic spectra were extracted and the kernel regression model between the Lick indices and the atmospheric physical parame-ters was established using the KPLSR method. Then the physical parameters of DR8 measured spectral data were computed by the kernel regression model for testing. The test results were compared with the atmospheric physical parameters provided by SEGUE SSPP and were good results. In addition, we added a signal-to-noise ratio (SNR) of 10, 20, 30, 40, 50, 70, 90 and 120 Gaussian white noise to the Kurucz spectra. And the resulting spectra of different SNR were used to test the impact of noise on the parameter measurement. The experimental results show that the kernel regression model is sensitive to noise, the higher the SNR of spectral data, the higher the prediction accuracy of the physical parameters. The method of KPLSR based on Lick line index has small amount of computation and fast training speed, which is suitable for measuring physical parameters of stellar at-mosphere.

[Searching for dwarf nova candidates with automatic methods in massive spectra].

In the present paper, an automatic and efficient method for searching for dwarf nova candidates is presented. The methods PCA (principal component analysis) and SVM (support vector machine) are applied in the newly released SDSS-DR9 spectra. The final dimensions of the feature space are determined by the identification accuracy of training samples with different dimensions constrained by SVM. The massive spectra are dimension reduced by PCA at first and classified by the best SVM clas sifier. The final less number of candidates can be identified manually. A total number of 276 dwarf nova candidates are selected by the method and 6 of them are new discoveries which prove that our approach to finding special celestial bodies in massive spectra data is feasible. The new discoveries of this paper are added in the current dwarf nova template library which can contribute to constructing a more accurate feature space. The method proposed in this paper can also be used for special objects searching in other sky survey telescopes like Guoshoujing (Large Sky Area Multi-Object Fiber Spectroscopic Telescope -LAMOST) telescope.

[The study of M dwarf spectral classification].

As the most common stars in the galaxy, M dwarfs can be used to trace the structure and evolution of the Milky Way. Besides, investigating M dwarfs is important for searching for habitability of extrasolar planets orbiting M dwarfs. Spectral classification of M dwarfs is a fundamental work. The authors used DR7 M dwarf sample of SLOAN to extract important features from the range of 600-900 nm by random forest method. Compared to the features used in Hammer Code, the authors added three new indices. Our test showed that the improved Hammer with new indices is more accurate. Our method has been applied to classify M dwarf spectra of LAMOST.

[Data mining for cataclysmic variables candidates in SDSS-DR8].

An automatic and efficient method for cataclysmic variables candidates is presented in this paper. The nonlinear locally linear embedding-LLE method is applied in the newly released SDSS-DR8 spectra. Spectra are dimension-reduced by LLE and classified by artificial neural network. The greatly reduced final candidates can be identified manually. 6 new CVs candidates were found in the experiment, and the compare between LLE with PCA shows the feasibility of nonlinear method in data mining in astronomical data.

[Comparison of different template matching methods of automatic stellar spectrum parameter measurement].

Template matching is one of the most commonly used methods of automatic stellar spectrum parameter measurement The present paper made comparisons among the three commonly used template matching methods K-Nearest Neighbor (KNN), the Chi-square minimization and cross correlation method. The Continuum normalization and flux normalization were made first. Then the three mentioned methods were compared in measurement results of stellar spectrum parameters. Experiments on SDSS DR8 large sample spectra showed that the cross correlation method had comparable advantages.

[Stellar spectrum parameter measurement based on line index by linear regression].

Using the Lick line index, according to the magnanimity characteristics of the spectrum an efficient algorithm of the atmospheric physical parameters measurement by the linear regression method from the point of view of statistical regression was designed. The linear regression was used to achieve the best regression effect by selecting the type of regression and the composition of line index. The formula obtained from the regression model makes the computation speed fast when applied to new data, and the clarity and ease of analysis processing can not be reached by other methods. The experimental results show that through the line index regression method to get the atmospheric physical parameters is feasible.

[An automated method to fit stellar continuum based on statistic windows].

A novel statistic window based method to fit stellar continuum is proposed. First a stellar spectrum is divided into a series of statistic windows in which a certain percent of flux points is selected according to S/N ratio; then low order polynomial iteration fitting is carried out based on the selected flux points to obtain the stellar continuum. Experimental results show that the continuum obtained by the proposed method is more close to the real continuum, compared to other existed methods. This method has a better practical applicability and robustness to all kinds of spectra (except M-type spectrum) in SDSS. It also works well for Guoshoujing Telescope (LAMOST) pilot survey spectra.

[A new automated method to identify emission line star from massive spectra].

Stellar spectra are characterized by obvious absorption lines or absorption bands, while those with emission lines are usually special stars such as cataclysmic variable stars (CVs), HerbigAe/Be etc. The further study of this kind of spectra is meaningful. The present paper proposed a new method to identify emission line stars (ELS) spectra automatically. After the continuum normalization is done for the original spectral flux, line detection is made by comparing the normalized flux with the mean and standard deviation of the flux in its neighbor region The results of the experiment on massive spectra from SDSS DR8 indicate that the method can identify ELS spectra completely and accurately. Since no complex transformation and computation are involved in this method, the identifying process is fast and it is ideal for the ELS detection in large sky survey projects like LAMOST and SDSS.

[Automatic stellar spectrum parameter measurement based on similarity measure].

In stellar spectral processing, template matching method can be used to obtain comparably ideal results of stellar spectrum parameters by using pattern recognition algorithms, without computing the spectral line-index. The present paper proposed a similarity based method to measure stellar spectrum parameters automatically. First, the continuum normalization was made, and then the similarities between the stellar spectrum and the template spectrum were compared to get more accurate stellar spectrum parameters. Experiments on ELODIE spectrum showed that this method is ideal in efficiently obtaining stellar spectrum parameters.