一、高定标精度下的测光红移测量

Photometric redshift measurements with high calibration precision

测量河外星系的距离是开展更多更有意义的科学研究的前提。除了利用那些与红移无关的标准距离示踪天体外，估计距离的最佳方法是通过红移，而红移只能通过昂贵且耗时的光谱分析进行精确测量。另一方面，通过测量星系不同波段流量得到的多波段颜色信息，获取星系光谱能量分布（SED）的稀疏采样，同样可以估算红移，这种方法称之为测光红移，是另一种相对经济的测量宇宙学距离的方法(Bilicki et al. 2018; Salvato et al. 2019)。

The distance of an extragalactic source needs to be carefully measured before any more meaningful physical quantities can be inferred. Except for those mostly local samples of redshift-independent “distance indicators”, the best way of estimating distance is via redshift, which could be measured precisely only from expensive and time-consuming spectroscopy. Photometric redshift, on the other hand, is another relatively cheap method to estimate cosmological distance, which is estimated by the colors of a galaxy derived from flux measurements in different bands (Bilicki et al. 2018; Salvato et al. 2019). The obtaining sparse sampling of the spectral energy distribution (SED) from photometry makes it possible to derive the redshift  in an imaging survey, but at the price of a lower redshift precision.

经过数十年的发展，人们提出了两种主要的方法（Brodwin et al.2006；Hildebrandt et al. 2010），即：（1）模板拟合；（2）机器学习（ML），并经过多次改进以提高测光红移测量的准确度和精确度。对于模板拟合方法，定义一组完备SED模板至关重要。这些模板可以从理论或实际观测中获得（例如，Fioc &Rocca Volmernage et al. 1997；Kinney et al.1996；Polletta et al. 2007；Conroy  2012）。从很大程度上来说，利用该方法所得的测光红移准确度非常依赖于模板类型及在颜色-红移空间中的完备性。对于机器学习方法，人们通常采用人工神经网络（Tagliaferri et al.2003）、增强决策树及回归树（Gerdes et al.2010）或遗传算法（Hogan et al.2015）等技术，基于光谱样本得到从测光数据到红移的映射关系。这些方法的关键在于建立足够具有代表性的训练集样本。整体来说，贝叶斯测光红移（BPZ，Benítez 2000）拟合方法和人工神经网络机器学习方法ANNz2（Sadeh et al.2016）是其中两种常用方法，已广泛应用于实际巡天数据（CFHTLenS、KiDS等）。

With decades of development, two main approaches (Brodwin et al. 2006; Hildebrandt et al. 2010), i.e., (i) template fitting; (2) machine-learning (ML) are proposed and have been through several revolutions to improve the accuracy and precision of the photometric redshift measurements. For template fitting approach, the most basic ingredient is the definition of a set of SED templates, which could be obtained either from theory or from observations (e.g., Fioc & Rocca-Volmernage et al. 1997; Kinney et al. 1996; Polletta et al. 2007; Conroy 2012). Therefore, the quality of photo-zs depends strongly on both the type of templates and their optimal coverage in the colour-redshift space. For ML approach, techniques such as artificial neural networks (Tagliaferri et al. 2003), boosted decision or regression trees (Gerdes et al. 2010), or genetic algorithms (Hogan et al. 2015), are calibrated (trained) on spec-z samples, to derive the mapping from photometry to redshifts. These methods are usually agnostic to any physics, and thus need well controlled and representative training sets to work properly. Generally speaking, two representative pipelines, i.e., Bayesian Photometric Redshift (BPZ, Benítez 2000) SED-fitting code and artificial neural network machine-learning approach ANNz2 (Sadeh et al. 2016), are commonly applied in real survey data (CFHTLenS, KiDS, etc.).

除了不同估计方法可能引入的影响因素以外，因定标精度有限导致的颜色不准确性是制约测光红移测量准确性和精确性的另一个重要系统误差。颜色的不准确性可以改变测量到的星系光谱能量分布，这不仅会使测光红移测量产生弥散，还会造成测量偏差。因此，我们期待，一个具有更高测光定标精度的大规模多波段巡天可以获得更好的测光红移测量。

Besides the possible effects introduced in different estimation methods, color inaccuracy from limited photometric calibration precision can be another important factor to inhibit the photo-z accuracy and precision. The color inaccuracy could  change the measured SED distribution of the galaxies, which will not only lead to  scatters but also cause biases in photo-z measurements. Therefore, large-scale multi-band surveys with higher photometric calibration precision could be crucial to make better photo-z measurements.  

Mephisto是国际上首台大视场、大口径、多通道高精度成像巡天望远镜，主镜口径1.6米，视场3.14平方度，首次通过创新性设计实现三通道分光，配备三台超大靶面拼接CCD相机，其独特优势在于可同时在3个波段（ugi或vrz）对同一天区进行高质量成像观测，获取天体高精度实时颜色信息（定标精度高达0.2-0.5%），录制宇宙天体运动和变化的彩色纪录片。我们期待，Mephisto提供的高质量颜色信息将可大幅提升测光红移的测量准确度和精确性。在本课题中，我们计划开展详细研究探讨定标精度对测光红移测量的影响。该研究不仅可以深入理解颜色准确性对测光红移测量的重要性，还将对Mephisto的定标精度提出更加具体的要求。

The Multi-channel Photometric Survey Telescope (Mephisto) has a 1.6m primary with a large field-of-view of 3.14 deg2 and is equipped with three CCD cameras, capable of simultaneously imaging the same patch of sky in three bands (ugi or vrz). Mephisto will yield real-time colors of astronomical objects with unprecedented accuracies (with ultra-high 0.2-0.5% calibration precision), and deliver for the first time a coloured documentary of our evolving universe. It is expected that the high quality color data provided by Mephisto will allow us to significantly improve the accuracy and precision of photo-z measurements. In the proposal, we propose to carry out detailed investigation on the issue that how the calibration precision influence the photo-z estimation. This study will not only study the importance of color accuracy in photo-z measurements but also put forward more specific requirements on calibration precision for Mephisto survey.

二、Mephisto巡天中恒星/星系分类精度的研究

Star-galaxy classification in the Mephisto survey

大视场巡天中，恒星/星系的分类对于研究宇宙学和银河系科学具有重要意义。传统方法利用天体的形态可以对亮源实现高精度的分类，然而，暗端的星系和恒星则会相互污染。为此，许多大视场巡天观测（如SDSS、Pan-STARRS等）融合了颜色信息以提高分类的完备性和纯度。除此之外，机器学习也已于最近几年发展成为一个重要的分类工具，并成功应用在许多巡天之中，如SDSS、CFHTLenS及DES等。

The classification of sources into stars and galaxies in wide field surveys is crucial for many areas of cosmology and Galactic science.  Conventionally, morphology-based methods can separate point sources (stars and quasars) from resolved sources (galaxies) with high accuracy for bright sources. At fainter magnitudes, however, unresolved galaxies can contaminate the point sources and noisy measurements of stars can also contaminate the galaxy sample. Therefore, many large multi-band imaging surveys, such as SDSS (Stoughton et al. 2002) and  Pan-STARRS (Saglia et al. 2012), have incorporated color information to improve the completeness and purity of the classification. Besides, machine learning (ML) has emerged as an important tool for classification in recent years and has been successfully applied on many surveys, such SDSS (Vasconcellos et al. 2011), CFHTLS (Kim et al. 2015), and DES (Sevilla-Noarbe et al. 2018).

相比于其他巡天1%的定标精度（如SDSS），Mephisto的高精度颜色定标精度有望对天体进行准确地分类，并提升我们对传统天体物理的理解。本研究拟针对Mephisto巡天，利用机器学习方法定量分析不同颜色定标精度对恒星/星系分类效率的影响。

Compared with other surveys with about one percent calibration precision (such as SDSS), it is expected that the improved color measurements of Mephisto will allow us to classify the objects with high precision and dramatically improve the understanding of traditional astrophysics. Therefore, we propose to study the capability of object classification in the Mephisto survey under its ultra-high calibration precision. In the proposal, we will focus on the performance of machine learning algorithms on the classification efficiency under different color accuracies.