天文光谱特征提取及其应用研究

With the recent implementation of some large-scale astronomical sky survey proposals (e.g. SDSS, 2dF and LAMOST), celestial spectra are becoming very abundant and rich. Therefore, the following problems are urgent to be solved: how to quickly estimate the parameters of the high-dimentional spectra and file them, how to discover the patterns under the enormous spectra (pattern exploring), how to retrieve them based on some specified contents in spectra, etc. One of the key procedures in solving the above problems is to extract the spectrum features and represent the spectra adequately and compactly. However, the traditional spectrum features are vulnerable to noise, calibration-distortion, spectrum complexity, redshift and its range, which result that the accuracy of processed result is unstable and that the labeled celestial spectra can not be effectively used in automatic spectra between different telescopes. Furthermore, the idea based on interest point has been applified sussesfully in computer vision community, and we proposed a novel model correspondence function(CF) for analyzing image interest point features.To deal with the above limitations of the traditional methods, this project focus on designing a novel feature extraction method for astronomical spectra based on the idea of image interest point by localizing the discriminating information，adaptively locating and rescaling it, discreterize and statistically quantizating the features. The research contents are as following: 1) Investigate how to detect, locate and represent the spectrum features robustly; 2) Based on the extraced features and the CF model, design a scheme to recognizing the Qso spectrm by utilizing a large amount of historical labeled spectra from different telescope, e.g. thousands of millions of them.

随着SDSS、2dF和LAMOST等大型天文观测项目的实施，天体光谱的获取速度和数据量急剧增加。如何快速、高质量地对光谱进行参数估计、分类归档、检索和物理规律探索成为有效利用这些海量、高维数据所急需解决的问题。其关键环节是光谱特征的提取，而传统方法对噪声、定标质量、光谱复杂度、红移及其变化范围等非常敏感，导致基于它的分析结果精度不稳定、并影响已标注光谱的跨望远镜使用。特征点思想已成功应用于计算机视觉领域，我们提出了特征点分析模型Correspondence Function(CF)，初步研究表明它可用于光谱自动分析。本项目旨在基于特征点思想探索一种新的光谱特征提取方法，通过对特征的局部化、自定位、自适应尺度估计，及其表征的离散化和统计量化克服上述因素的不良影响。内容包括：光谱特征的检测、定位及其稳健表征方法研究；基于该特征和大规模跨望远镜历史标注光谱，研究CF模型在类星体识别中的应用。

随着SDSS、LAMOST、Gaia和FAST等大型天文观测项目的实施，天体数据的获取速度和数据量雪崩式增长，天文学成文一个数据密集型科学和一个典型的大数据应用领域。本项目研究了天文光谱数据中特征的检测、描述与应用问题, 提出了谱线信息组合特征的自动计算发现方案，而文献中已有研究则是通过人工方式经验探索。具体内容包括：1) 在特征的检测方面，给出了基于稀疏最优化理论、非最大抑制思想的天文光谱特征发现与定位方法；2)特征的稳健描述方面，提出了基于局部积分、卷积与池化、小波变换，栈式自编码神经网络的光谱特征描述方案；3)在稳健特征的天体目标识别应用研究方面，探索了类星体光谱、星系光谱、恒星光谱分类(相关文献中亦有称为恒星光谱的参数化)、和白矮主序双星光谱的识别等。上述研究内容均做了一定研究、并有相关成果，研究目标基本按计划完成。在国内外权威期刊The Astrophysical Journal Supplement Series(SCI收录期刊，影响因子11.257), The Astrophysical Journal(SCI收录期刊，影响因子5.909)，Monthly Notices of the Royal Astronomical Society (SCI收录期刊，影响因子4.952)，Astronomy and Astrophysics (SCI收录期刊，影响因子5.185)等权威期刊和学术会议发表研究论文13篇。该研究结果对于天文大数据的处理、模型探索、因果关系的研究有重要的参考价值。