基于SE-Inception-v3的星系形态分类模型

随着天文探测技术的快速发展, 海量的星系图像数据不断产生, 能够及时高效地对星系图像进行形态分类对研究星系的形成与演化至关重要. 针对传统的星系形态分类模型特征选择困难、分类速度慢、准确率受限等难题, 提出一种以Inception-v3神经网络为主干结构, 融合压缩激励(Squeeze and Excitation Network, SE)通道注意力机制的星系形态分类模型. 该模型在斯隆数字巡天(Sloan Digital Sky Survey, SDSS)样本的测试集准确率高达99.37%. 旋涡星系、圆形星系、中间星系、雪茄状星系与侧向星系的F1值分别为99.33%、99.58%、99.33%、99.41%与99.16%. 该模型与Inception-v3、MobileNet (Mobile Neural Network)和ResNet (Residual Neural Network)网络模型相比, SE-Inception-v3宽度和深度优势表现出更强的特征提取能力, 可以高效识别不同形态的星系, 为未来大型巡天计划的大规模星系形态分类问题提供了一种新方法.     

基于Stacking集成学习的恒星/星系分类研究

机器学习在当今诸多领域已经取得了巨大的成功,但是机器学习的预测效果往往依赖于具体问题.集成学习通过综合多个基分类器来预测结果,因此,其适应各种场景的能力较强,分类准确率较高.基于斯隆数字巡天(Sloan Digital Sky Survey,SDSS)计划恒星/星系中最暗源星等集分类正确率低的问题,提出一种基于Stacking集成学习的恒星/星系分类算法.从SDSS-DR7(SDSS Data Release 7)中获取完整的测光数据集,并根据星等值划分为亮源星等集、暗源星等集和最暗源星等集.仅针对分类较为复杂且困难的最暗源星等集展开分类研究.首先,对最暗源星等集使用10折嵌套交叉验证,然后使用支持向量机(Support Vector Machine,SVM)、随机森林(Random Forest,RF)、XGBoost(eXtreme Gradient Boosting)等算法建立基分类器模型;使用梯度提升树(Gradient Boosting Decision Tree,GBDT)作为元分类器模型.最后,使用基于星系的分类正确率等指标,与功能树(Function Tree,FT)、SVM、RF、GBDT、XGBoost、堆叠降噪自编码(Stacked Denoising AutoEncoders,SDAE)、深度置信网络(Deep Belief Network,DBN)、深度感知决策树(Deep Perception Decision Tree,DPDT)等模型进行分类结果对比分析.实验结果表明,Stacking集成学习模型在最暗源星等集分类中要比FT算法的星系分类正确率提高了将近10%.同其他传统的机器学习算法、较强的提升算法、深度学习算法相比,Stacking集成学习模型也有较大的提升.     

基于K近邻方法的窄线与宽线活动星系核的自动光谱分类

对于美国芝加哥大学等6个组织的斯隆数字化巡天观测(SDSS)的一批低红移活动星系核(AGN)光谱数据,针对宽线AGNs和窄线AGNs发射线的不同特征,在静止系的光谱上截取有效波段范围,采用自动分类的K近邻方法,对其进行分类．宽线和窄线AGNs光谱的主要区别在于Hβ、[OⅢ]、Ha和[NⅡ]等发射线的幅度和半高全宽(FWHM)的大小,所以截取这些发射线所在的波段进行单独或组合的分类实验,实验证明,单独采用以Hα和[NⅡ]发射线为主的波段,分类效果最好,且对于训练样本数和测试样本数分别为1000和3313条的AGNs光谱的单次分类速度可达32．89秒．在充分利用光谱的典型特征的情况下,自动分类方法也可有效地应用于活动星系核的分类,为传统的通过计算发射线的FWHM值或发射线强比对大型光谱巡天所产生的庞大数据库进行分类提供了一种快速直接的分类方法．     

基于深度残差网络的星系形态分类

星系形态与星系的形成和演化有着密切的联系,因此星系形态分类(galaxy morphology classification)成为研究不同星系物理特征的重要过程之一。斯隆数字巡天(Sloan Digital Sky Survey, SDSS)等大型巡天计划产生的海量星系图像数据对星系形态的准确、实时分类提出了新的挑战,而深度学习(deep learning)算法能有效应对这类海量星系图片的自动分类考验。面向星系形态分类问题提出了一种改进的深度残差网络(residual network, ResNet),即ResNet-26模型。该模型对残差单元进行改进,减少了网络深度,并增加了网络宽度,实现了对星系形态特征的自动提取、识别和分类。实验结果表明,与Dieleman和ResNet-50等其他流行的卷积神经网络(convolution neural network, CNN)模型相比,ResNet-26模型具有更优的分类性能,可应用于未来大型巡天计划的大规模星系形态分类系统。     

基于XGBoost算法的恒星/星系分类研究

机器学习在当今的诸多领域已经取得了巨大的成功.尤其是提升算法.提升算法适应各种场景的能力较强、准确率较高,已经在多个领域发挥巨大的作用.但是提升算法在天文学中的应用却极为少见.为解决斯隆数字巡天(Sloan Digital Sky Survey,SDSS)数据中恒星/星系暗源集分类正确率低的问题,引入了机器学习中较新的研究成果–XGBoost (eXtreme Gradient Boosting).从SDSS-DR7 (SDSS Data Release 7)中获取完整的测光数据集,并根据星等值划分为亮源集和暗源集.首先,分别对亮源集和暗源集使用十折交叉验证法,同时运用XGBoost算法建立恒星/星系分类模型;然后,运用栅格搜索等方法调优XGBoost参数;最后,基于星系的分类正确率等指标,与功能树(Function Tree, FT)、Adaboost (Adaptive boosting)、随机森林(Random Forest, RF)、梯度提升决策树(Gradient Boosting Decision Tree, GBDT)、堆叠降噪自编码(Stacked Denoising AutoEncoders, SDAE)、深度置信网络(Deep Belief Network, DBN)等模型进行对比并分析结果.实验结果表明:XGBoost在暗源分类中要比功能树算法的星系分类正确率提高了将近10%,在暗源集的最暗星等中比功能树提高了将近5%.同其他传统的机器学习算法和深度神经网络相比, XGBoost也有不同程度的提升.     

EfficientNetV2-S-Triplet7: 一种改进的星系newline 形态学分类算法

星系的形态与星系的形成和演化息息相关, 其形态学分类是星系天文学后续研究的重要一环. 当前海量天文观测数据的出现使得天文数据自动分析方法越来越得到重视, 针对此问题, 利用先进的深度学习骨干网络EfficientNetV2, 分析不同的注意力机制类型和使用节点对网络性能的影响, 构建了一种命名为EfficientNetV2-S-Triplet7 (即在EfficientNetV2-S stage7的$1\times1$卷积层后加入Triplet模块)的改进算法模型来实现星系形态学的自动分类. 使用第二期星系动物园(Galaxy Zoo 2, GZ2)中超过24万张的测光图像作为初始数据进行实验测试. 在对数据进行预处理时采取了尺寸抖动、翻转、色彩畸变等图像增强手段来解决图像数量的不平衡问题. 在同一系列经典和前沿的深度学习算法模型AlexNet、ResNet-34、MobileNetV2、RegNet进行对比实验后, 得出EfficientNetV2-S-Triplet7算法在分类准确率、查全率和F1分数等指标上具有最好的测试结果. 在9375张测试图像中的3项指标值分别可达到89.03%、90.21%、89.93%, 查准率达到89.69%, 在其他模型中排在第3位. 该结果表明将EfficientNetV2-S-Triplet7算法应用于大规模星系数据的形态学分类任务中有很好的效果.     

基于NPSVM模型的超新星识别方法

巡天观测与高能物理、黑洞天文等领域均有密切的联系.基于星系-超新星二分类问题,研究光谱数据预处理,结合余弦相似度改善PCA(Principal Component Analysis)光谱分解特征提取方法,用SDSS(the Sloan Digital Sky Survey)、WISeREP(the Weizmann Interactive Supernova data REPository)组成的5620条光谱数据集训练支持向量机,可以得到0.498%泛化误差的识别模型和新样本分类概率.使用Neyman-Pearson决策方法建立NPSVM(Neyman-Pearson Support Vector Machine)模型可进一步降低超新星的漏判率.     

基于EfficientNet的星系形态分类研究

星系的结构和形态能够反映星系自身的物理性质,其形态的分类是后续分析研究的一个重要环节.EfficientNet模型使用复合系数对深度网络模型的深度、宽度、输入图像分辨率进行更加结构化的统一缩放,是一种新的深度网络优化扩展方法.将该模型应用于星系数据形态的分类研究中,结果表明基于EfficientNetB5模型的平均准确率、精确率、召回率以及F1分数(精确率与召回率的调和平均数)都在96.6%以上,与残差网络(Residual network, ResNet)中ResNet-26模型的分类结果相比有较大的提升.实验结果证明EfficientNet的深度网络优化扩展方法可行且有效,可应用于星系的形态分类.     

基于深层残差网络的脉冲星候选体分类方法研究

随着下一代射电天文望远镜的不断改进和发展,脉冲星巡天观测将发现数百万个脉冲星候选体,这给脉冲星的识别和新脉冲星的发现带来了巨大挑战,迅速发展的人工智能技术可用于脉冲星识别.使用Parkes望远镜的脉冲星数据集(The High Time Resolution Universe Survey,HTRUS),设计了一个14层深的残差网络(Residual Network,ResNet)进行脉冲星候选体分类.在HTRUS数据样本中,存在非脉冲星候选体(负样本)的数目远远大于脉冲星候选体(正样本)数目的样本非均衡问题,容易产生模型误判.通过使用过采样技术对训练集中的正样本进行数据增强,并调整正负样本的比例,解决了正负样本非均衡问题.训练过程中,使用5折交叉验证来调节超参数,最终构建出模型.测试结果表明,该模型能够取得较高的精确度(Precision)和召回率(Recall),分别为98%和100%,F1分数(F1-score)能够达到99%,每个样本检测完成只需要7 ms,为未来脉冲星大数据分析提供了一个可行的办法.     

同等质量下盘状星系的颜色-尺度关系

挑选Sloan数字巡天第7次释放数据(SDSS DR7)的主星系样本中近邻的、面向的盘状星系作为星系样本,统计研究了在恒星总质量相等的情况下盘状星系的颜色和尺度之间的相关性,并对相关性的真实性进行了检验.发现对于同等质量的盘状星系,u-r颜色与尺度相关性很弱,而g-r、r-i、r-z颜色与尺度负相关,即星系的尺度越大,颜色越蓝.该结果意味着盘状星系的质量分布对其恒星形成历史影响很大,物质分布越延展的星系,其演化越慢.     

修正的极大似然法拟合幂律频数分布

针对包含饱和样本数据的频数幂律分布拟合,提出一个新的幂律分布指数的极大似然估计方法的修正公式.对比研究显示,修正公式适用于剔除异常饱和值的幂律频数拟合.如果不剔除饱和值,幂律指数的估计只能使用修正前的公式,其误差随幂律指数变化,指数较小时误差较大.由此建议,对于包含饱和样本的频数分布拟合,首先剔除异常的饱和值,然后对剩余不含饱和值的子集使用修正公式进行参数估计.     

Cosmological systematics beyond nuisance parameters: form-filling functions

In the absence of any compelling physical model, cosmological systematics are often misrepresented as statistical effects and the approach of marginalizing over extra nuisance systematic parameters is used to gauge the effect of the systematic. In this article, we argue that such an approach is risky at best since the key choice of function can have a large effect on the resultant cosmological errors.
As an alternative we present a functional form-filling technique in which an unknown, residual, systematic is treated as such. Since the underlying function is unknown, we evaluate the effect of every functional form allowed by the information available (either a hard boundary or some data). Using a simple toy model, we introduce the formalism of functional form filling. We show that parameter errors can be dramatically affected by the choice of function in the case of marginalizing over a systematic, but that in contrast the functional form-filling approach is independent of the choice of basis set.
We then apply the technique to cosmic shear shape measurement systematics and show that a shear calibration bias of  | m ( z )| ≲ 10⁻³ (1 + z )^0.7  is required for a future all-sky photometric survey to yield unbiased cosmological parameter constraints to per cent accuracy.
A module associated with the work in this paper is available through the open source icosmo code available at http://www.icosmo.org .     

Substructure recovery by three-dimensional discrete wavelet transforms

We present and discuss a method to identify substructures in combined angular-redshift samples of galaxies within clusters. The method relies on the use of the discrete wavelet transform (hereafter DWT) and has already been applied to the analysis of the Coma cluster. The main new ingredient of our method with respect to previous studies lies in the fact that we make use of a 3D data set rather than a 2D one. We test the method on mock cluster catalogues with spatially localized substructures and on a N -body simulation. Our main conclusion is that our method is able to identify the existing substructures provided that: (a) the subclumps are detached in part or all of the phase space, (b) one has a statistically significant number of redshifts, increasing as the distance decreases due to redshift distortions; (c) one knows a priori the scale on which substructures are to be expected. We have found that to allow an accurate recovery we must have both a significant number of galaxies (≈200 for clusters at z ≥0.4 or about 800 at z ≤0.4) and a limiting magnitude for completeness m _B =16.
The only true limitation to our method seems to be the necessity of knowing a priori the scale on which the substructure is to be found. This is an intrinsic drawback of the method and no improvement in numerical codes based on this technique could make up for it.     

Accurate estimators of power spectra in N-body simulations

A method to rapidly estimate the Fourier power spectrum of a point distribution is presented. This method relies on a Taylor expansion of the trigonometric functions. It yields the Fourier modes from a number of fast Fourier transforms (FFTs), which is controlled by the order N of the expansion and by the dimension D of the system. In three dimensions, for the practical value   N = 3  , the number of FFTs required is 20.
We apply the method to the measurement of the power spectrum of a periodic point distribution that is a local Poisson realization of an underlying stationary field. We derive an explicit analytic expression for the spectrum, which allows us to quantify – and correct for – the biases induced by discreteness and by the truncation of the Taylor expansion, and to bound the unknown effects of aliasing of the power spectrum. We show that these aliasing effects decrease rapidly with the order N . For   N = 3  , they are expected to be, respectively, smaller than  ∼10⁻⁴  and 0.02 at half the Nyquist frequency and at the Nyquist frequency of the grid used to perform the FFTs. The only remaining significant source of errors is reduced to the unavoidable cosmic/sample variance due to the finite size of the sample.
The analytical calculations are successfully checked against a cosmological N -body experiment. We also consider the initial conditions of this simulation, which correspond to a perturbed grid. This allows us to test a case where the local Poisson assumption is incorrect. Even in that extreme situation, the third-order Fourier–Taylor estimator behaves well, with aliasing effects restrained to at most the per cent level at half the Nyquist frequency.
We also show how to reach arbitrarily large dynamic range in Fourier space (i.e. high wavenumber), while keeping statistical errors in control, by appropriately 'folding' the particle distribution.     

Detrending time series for astronomical variability surveys

We present a detrending algorithm for the removal of trends in time series. Trends in time series could be caused by various systematic and random noise sources such as cloud passages, changes of airmass, telescope vibration, CCD noise or defects of photometry. Those trends undermine the intrinsic signals of stars and should be removed. We determine the trends from subsets of stars that are highly correlated among themselves. These subsets are selected based on a hierarchical tree clustering algorithm. A bottom-up merging algorithm based on the departure from normal distribution in the correlation is developed to identify subsets, which we call clusters. After identification of clusters, we determine a trend per cluster by weighted sum of normalized light curves. We then use quadratic programming to detrend all individual light curves based on these determined trends. Experimental results with synthetic light curves containing artificial trends and events are presented. Results from other detrending methods are also compared. The developed algorithm can be applied to time series for trend removal in both narrow and wide field astronomy.     

The analysis of irregularly observed stochastic astronomical time-series – I. Basics of linear stochastic differential equations

The theory of low-order linear stochastic differential equations is reviewed. Solutions to these equations give the continuous time analogues of discrete time autoregressive time-series. Explicit forms for the power spectra and covariance functions of first- and second-order forms are given. A conceptually simple method is described for fitting continuous time autoregressive models to data. Formulae giving the standard errors of the parameter estimates are derived. Simulated data are used to verify the performance of the methods. Irregularly spaced observations of the two hydrogen-deficient stars FQ Aqr and NO Ser are analysed. In the case of FQ Aqr the best-fitting model is of second order, and describes a quasi-periodicity of about 20 d with an e-folding time of 3.7 d. The NO Ser data are best fitted by a first-order model with an e-folding time of 7.2 d.