Automated classification of stellar spectra — II. Two-dimensional classification with neural networks and principal components analysis

We investigate the application of neural networks to the automation of MK spectral classification. The data set for this project consists of a set of over 5000 optical (3800–5200 Å) spectra obtained from objective prism plates from the Michigan Spectral Survey. These spectra, along with their two-dimensional MK classifications listed in the Michigan Henry Draper Catalogue, were used to develop supervised neural network classifiers. We show that neural networks can give accurate spectral type classifications (σ₆₈= 0.82 subtypes, σ_rms= 1.09 subtypes) across the full range of spectral types present in the data set (B2–M7). We show also that the networks yield correct luminosity classes for over 95 per cent of both dwarfs and giants with a high degree of confidence.   Stellar spectra generally contain a large amount of redundant information. We investigate the application of principal components analysis (PCA) to the optimal compression of spectra. We show that PCA can compress the spectra by a factor of over 30 while retaining essentially all of the useful information in the data set. Furthermore, it is shown that this compression optimally removes noise and can be used to identify unusual spectra.   This paper is a continuation of the work carried out by von Hippel et al. (Paper I).     

Interpreting Yohkoh hard and soft X-ray flare observations

A simple model is presented to account for theYohkoh flare observations of Feldmanet al. (1994), and Masuda (1994). Electrons accelerated by the flare are assumed to encounter the dense, small regions observed by Feldmanet al. at the tops of impulsively flaring coronal magnetic loops. The values of electron density and volume inferred by Feldmanet al. imply that these dense regions present an intermediate thick-thin target to the energised electrons. Specifically, they present a thick (thin) target to electrons with energy much less (greater) thanE _c , where 15 keV <E _c < 40 keV. The electrons are either stopped at the loop top or precipitate down the field lines of the loop to the footpoints. Collisional losses of the electrons at the loop top produce the heating observed by Feldmanet al. and also some hard X-rays. It is argued that this is the mechanism for the loop-top hard X-ray sources observed in limb flares by Masuda. Adopting a simple model for the energy losses of electrons traversing the dense region and the ambient loop plasma, hard X-ray spectra are derived for the loop-top source, the footpoint sources and the region between the loop top and footpoints. These spectra are compared with the observations of Masuda. The model spectra are found to qualitatively agree with the data, and in particular account for the observed steepening of the loop-top and footpoint spectra between 14 and 53 keV and the relative brightnesses of the loop-top and footpoint sources.     

Stellar Spectral Classification using Principal Component Analysis and Artificial Neural Networks

A fast and robust method of classifying a library of optical stellar spectra for O to M type stars is presented. The method employs, as tools: (1) principal component analysis (PCA) for reducing the dimensionality of the data and (2) multilayer back propagation network (MBPN) based artificial neural network (ANN) scheme to automate the process of classification. We are able to reduce the dimensionality of the original spectral data to very few components by using PCA and are able to successfully reconstruct the original spectra. A number of NN architectures are used to classify the library of test spectra. Performance of ANN with this reduced dimension shows that the library can be classified to accuracies similar to those achieved by Gulati et al. but with less computational load. Furthermore, the data compression is so efficient that the NN scheme successfully classifies to the desired accuracy for a wide range of architectures. The procedure will greatly improve our capabilities in handling and analysing large spectral data bases of the future.     

Excess B‐modes extracted from the Planck polarization maps

One of the main obstacles for extracting the Cosmic Microwave Background (CMB) from mm/submm observations is the pollution from the main Galactic components: synchrotron, free‐free and thermal dust emission. The feasibility of using simple neural networks to extract CMB has been demonstrated on both temperature and polarization data obtained by the WMAP satellite. The main goal of this paper is to demonstrate the feasibility of neural networks for extracting the CMB signal from the Planck polarization data with high precision. Both auto‐correlation and cross‐correlation power spectra within a mask covering about 63 % of the sky have been used together with a “high pass filter” in order to minimize the influence of the remaining systematic errors in the Planck Q and U maps. Using the Planck 2015 released polarization maps, a BB power spectrum have been extracted by Multilayer Perceptron neural networks. This spectrum contains a bright feature with signal to noise ratios 4.5 within 200 ≪ l ≪ 250. The spectrum is significantly brighter than the BICEP2 2015 spectrum, with a spectral behaviour quite different from the “canonical” models (weak lensing plus B‐modes spectra with different tensor to scalar ratios). The feasibility of the neural network to remove the residual systematics from the available Planck polarization data to a high level has been demonstrated. (© 2016 WILEY‐VCH Verlag GmbH & Co. KGaA, Weinheim)     

Towards a new galaxy template library for multi-colour classification

The COMBO-17 survey cite(Wolf et al., 2002) contains ≈ 40000 galaxies down to R=24 mag on an area of one square degree, obtained with the wide field imager at the 2.2 m telescope at La Silla. A multi-colour classification on the basis of 5 broadband and 12 medium band filters (=17 bands) yields accurate redshifts (σ _z ≤ 0.01 at the bright end up to σ _z ≈ 0.1 at the faint end) and spectral energy distribution types (SEDs) when using observed galaxy templates from (Kinney et al.,1996). However, there is an obvious weakness in this classification scheme: The relation between star formation history and SED remains unclear. It is therefore impossible to draw firm conclusions about the age of the underlying stellar population and the expected aging between z ≈ 1 and z ≈ 0 can not be quantified. We will present first results of our attempt to replace the observed templates with model spectra from the PEGASE code (Fioc and Rocca-Volmerange, 1997), in order to get a better handle on the star formation history. This revised version was published online in August 2006 with corrections to the Cover Date.     

Spallation recoil and age of presolar grains in meteorites

Abstract— We have determined the recoil losses from silicon carbide (SiC) grain‐size fractions of spallation Ne produced by irradiation with 1.6 GeV protons. During the irradiation, the SiC grains were dispersed in paraffin wax in order to avoid reimplantation into neighboring grains. Analysis for spallogenic ²¹Ne of grain‐size separates in the size range 0.3 to 6 μm and comparison with the ²²Na activity of the SiC + paraffin mixture indicates an effective recoil range of 2–3 μm with no apparent effect from acid treatments, which are routinely used in the isolation of meteoritic SiC grains. Our results indicate that the majority of presolar SiC grains in primitive meteorites, which are micrometer‐sized, will have lost essentially all spallogenic Ne produced by cosmic‐ray interaction in the interstellar medium. This argues against the validity of previously published presolar ages of Murchison SiC (～10 to ～130 Ma, increasing with grain size; Lewis et al., 1994), where recoil losses had been based on calculated recoil energies. It is argued that the observed variations in meteoritic SiC grain‐size fractions of ²¹Ne/²²Ne ratios are more likely due to the effects of nucleosynthesis in the He‐burning shell of the parent AGB stars which imposes new boundary conditions on nuclear parameters and stellar models. It is suggested that spallation‐Xe produced on the abundant Ba and REE in presolar SiC, rather than spallogenic Ne, may be a promising approach to the presolar age problem. There is a hint in the currently available Xe data (Lewis et al., 1994) that the large (>1 μm) grains may be younger than the smaller (<1 μm) grains. The retention of spallogenic ²¹Ne produced by the bombardment of SiC grains of different grain sizes with 1.6 GeV protons, avoiding reimplantation into neighboring grains by dispersing the SiC grains in paraffin wax, has been derived from a comparison of mass spectrometrically determined ²¹Ne, retained in the grains, with the ²²Na activity of the grains‐plus‐paraffin mixture. Compared to estimates of retention used in previous attempts to determine presolar ages for SiC (Tang and Anders, 1988b; Lewis et al., 1990, 1994), the results indicate significantly lower values. They do, however, agree with retention as expected from previous measurements of recoil ranges in similar systems (Nyquist et al., 1973; Steinberg and Winsberg, 1974). The prime reason for the discrepancy must lie in the energy of the recoiling nuclei entering in the calculation of retention by Tang and Anders (1988b), which is based on considerations by Ray and Völk (1983). Based on the results, it appears questionable that spallation contributes significantly to the observed variations of ²¹Ne/²²Ne ratios among various SiC grain‐size separates (Lewis et al., 1994). We rather suggest that the variations, just as it has been observed for Kr and Ba already (Lewis et al., 1994; Prombo et al., 1993), have a nucleosynthetic origin. Confirmation needs input of improved nuclear data and stellar models into new network calculations of the nucleosynthesis in AGB stars of elements in the Ne region. Finally we argue that, to determine presolar system irradiation effects, spallation Xe is more favorable than is Ne, primarily because of smaller recoil losses for Xe. Although preliminary estimates hint at the possibility that the larger (>1 μm) grains are younger than the smaller (<1 μm) ones, the major uncertainty for a quantitative evaluation lies in the exact composition of the Xe‐N component thought to originate from the envelope of the SiC grains' parent stars.     

基于卷积神经网络的恒星光谱型和光度型的分类模型

恒星光谱分类是天文学中一个重要的研究问题.对于已经采集到的海量高维恒星光谱数据的分类,采用模式匹配方法对光谱型分类较为成功,但其缺点在于标准恒星模版之间的差异性在匹配实际观测数据中不能体现出来,尤其是当需要进行光谱型和光度型的二元分类时模版匹配法往往会失败.而采用谱线特征测量的光度型分类强烈地依赖谱线拟合的准确性.为了解决二元分类的问题,介绍了一种基于卷积神经网络的恒星光谱型和光度型分类模型(Classification model of Stellar Spectral type and Luminosity type based on Convolution Neural Network, CSSL CNN).这一模型使用卷积神经网络来提取光谱的特征,通过注意力模块学习到了重要的光谱特征,借助池化操作降低了光谱的维度并压缩了模型参数的数量,使用全连接层来学习特征并对恒星光谱进行分类.实验中使用了大天区面积多目标光纤光谱天文望远镜(Large Sky Area Multi-Object Fiber Spectroscopy Telescope, LAMOST)公开数据集Data Release 5 (DR5,用了其中71282条恒星光谱数据,每条光谱包含了3000多维的特征)对该模型的性能进行验证与评估.实验结果表明,基于卷积神经网络的模型在恒星的光谱型分类上准确率达到92.04%,而基于深度神经网络的模型(Celestial bodies Spectral Classification Model, CSC Model)只有87.54%的准确率; CSSL CNN在恒星的光谱型和光度型二元分类上准确率达到83.91%,而模式匹配方法MKCLASS仅有38.38%的准确率且效率较低.     

基于深度学习的LAMOST星系星族参数测量

星系的光谱包含其内部恒星的年龄和金属丰度等信息, 从观测光谱数据中测量这些信息对于深入了解星系的形成和演化至关重要. LAMOST (Large Sky Area Multi-Object Fiber Spectroscopic Telescope)巡天发布了大量的星系光谱, 这些高维光谱与它们的物理参数之间存在着高度的非线性关系. 而深度学习适合于处理多维、海量的非线性数据, 因此基于深度学习技术构建了一个8个卷积层$+$4个池化层$+$1个全连接层的卷积神经网络, 对LAMOST Data Release 7 (DR7)星系的年龄和金属丰度进行自动估计. 实验结果表明, 使用卷积神经网络通过星系光谱预测的星族参数与传统方法基本一致, 误差在0.18dex以内, 并且随着光谱信噪比的增大, 预测误差越来越小. 实验还对比了卷积神经网络与随机森林回归模型、深度神经网络的参数测量结果, 结果表明卷积神经网络的结果优于其他两种回归模型.     

Fullerenes, fulleranes and polycyclic aromatic hydrocarbons in the Allende meteorite

In this paper, we confirm our earlier observations of fullerenes (C60 and C70) in the Allende meteorite (Becker et al., 1994a, 1995). Fullerene C60 was also detected in two separate C-rich (approximately 0.5-1.0%) dark inclusions (Heymann et al., 1987) that were hand picked from the Allende sample. The amounts of C60 detected were approximately 5 and approximately 10 ppb, respectively, which is considerably less than what was detected in the Allende 15/21 sample (approximately 100 ppb; Becker et al., 1994a, 1995). This suggests that fullerenes are heterogeneously distributed in the meteorite. In addition, we present evidence for fulleranes, (C60Hx), detected in separate samples by laser desorption (reflectron) time-of-flight (TOF) mass spectrometry (LDMS). The LDMS spectra for the Allende extracts were remarkably similar to the spectra generated for the synthetic fullerane mixtures. Several fullerane products were synthesized using a Rh catalyst (Becker et al., 1993a) and separated using high-performance liquid chromatography (HPLC). Polycyclic aromatic hydrocarbons (PAHs) were also observed ppm levels) that included benzofluoranthene and corannulene, a cup-shaped molecule that has been proposed as a precursor molecule to the formation of fullerenes in the gas phase (Pope et al., 1993).     

Automatic Classification of Subdwarf Spectra using a Neural Network

We apply a multilayer feed-forward back propagation artificial neural network to a sample of 380 subdwarf spectra classified by Drilling et al. (Drilling, J.S., Moehler, S., Jeffery, C.S., Heber, U., and Napiwotzki, R.: in press in: R. Gray (ed.), Probing the Personalities of Stars and Galaxies), showing that it is possible to use this technique on large sets of spectra and obtain classifications in good agreement with the standard. We briefly investigate the impact of training set size, showing that large training sets do not necessarily perform significantly better than small sets.     

Short Gamma-Ray Bursts observed by PHEBUS

Statistical studies of Gamma-Ray Burst (GRB) properties have recently led to the discovery of a subclass within the population of classical events (Dezalayet al. 1992, Kouveliotouet al. 1993). Bursts belonging to this subclass are characterized by short durations, typically less than 2 seconds, and harder spectra on average. Using the PHEBUS GRB data set, we analyse the distributions of peak intensity, hardness ratio, and duration of the two subclasses. We also compare the sum spectra obtained with the brightest events to determine the ratio of total energies observed for each population.     

基于卷积神经网络的快速射电暴候选体分类

针对目前从海量的快速射电暴(Fast Radio Burst, FRB)候选体中人工筛选FRB事件难以为继的现状,提出了一种基于卷积神经网络(Convolutional Neural Networks, CNN)的FRB候选体分类方法.首先,通过真实的观测数据和仿真FRB组成训练和测试样本集.其次,建立了二输入的深度卷积神经网络模型,并对其进行训练、测试和优化,获取FRB候选体分类器.最后,利用来自脉冲星的单脉冲数据对该分类器的有效性和性能进行了验证.实验结果表明,该方法可以快速从大量候选体中准确识别出单脉冲事件,极大地提高了FRB候选体的处理速率和效率.     

Parameterizing Stellar Spectra Using Deep Neural Networks

Large-scale sky surveys are observing massive amounts of stellar spectra.The large number of stellar spectra makes it necessary to automatically parameterize spectral data,which in turn helps in statistically exploring properties related to the atmospheric parameters.This work focuses on designing an automatic scheme to estimate effective temperature(T_(eff)),surface gravity(log g) and metallicity[Fe/H] from stellar spectra.A scheme based on three deep neural networks(DNNs) is proposed.This scheme consists of the following three procedures:first,the configuration of a DNN is initialized using a series of autoencoder neural networks;second,the DNN is fine-tuned using a gradient descent scheme;third,three atmospheric parameters T_(eff),log g and [Fe/H] are estimated using the computed DNNs.The constructed DNN is a neural network with six layers(one input layer,one output layer and four hidden layers),for which the number of nodes in the six layers are 3821,1000,500,100,30 and 1,respectively.This proposed scheme was tested on both real spectra and theoretical spectra from Kurucz's new opacity distribution function models.Test errors are measured with mean absolute errors(MAEs).The errors on real spectra from the Sloan Digital Sky Survey(SDSS) are 0.1477,0.0048 and 0.1129 dex for log g,log T_(eff) and [Fe/H](64.85 K for T_(eff)),respectively.Regarding theoretical spectra from Kurucz's new opacity distribution function models,the MAE of the test errors are 0.0182,0.0011 and 0.0112 dex for log g,log T_(eff) and [Fe/H](14.90 K for T_(eff)),respectively.     

Feature relevance in morphological galaxy classification

We investigate the utility of a variety of features in performing morphological galaxy classification using back-propagation neural network classifiers based on a sample of 805 galaxies classified by Naim et al. We derive a total of 22 features from each galaxy image and use these as inputs to a neural network trained using back-propagation. The morphological types are subdivided into two to seven groups, and the relevance of each of the features is examined for each grouping. We use the magnitude of the regularization parameter for each input to determine whether a feature can be eliminated. We then prune the input features of the network, typically down to four features. We examine a number of methods of assessing the performance of the network and determine which works best for our task.     

CCD photometry of 6 near-Earth asteroids

Photometric observations of 11 near-Earth asteroids were made within a regular NEA CCD photometric programme at Ondejov Observatory in the first half of 1994. This paper shows obtained R lightcurves, V-R, R-I, and B-V color indices, and rotation preriods for 6 of them. Among the presented results, the most interesting are those for (4954) Eric, for which we obtained several high-quality lightcurves and which seems to indicate a surface heterogeneity, (1864) Daedalus, for which interesting comparison with older observations by Gehrelset al. (1971) can be made, and 1993 VW, for which the unusual color characteristics were observed. Two (1994 AW₁ and 1994 GY) of the other three observed objects have fast rotations with periods of about 2.5 hours and relatively low amplitudes. Subsequent analysis of additional lightcurve data for 1994 AW₁ has revealed a presence of two periods in its lightcurve (Pravecet al. 1995). The last object presented here is 1994 JF₁, for which only lower limits on period and amplitude were determined.     

Infrared,ultraviolet, and electron paramagnetic resonance measurements on presolar diamonds: Implications for optical features and origin

Abstract— Infrared (IR) and ultraviolet (UV) absorption spectra were obtained for diamonds from the Allende and Murchison meteorites. In addition, and for the first time, electron paramagnetic resonance spectra were measured. The IR and UV data confirm the suspicion of Russell et al. (1996) that N in presolar diamonds predominantly appears in the form of dispersed N atoms, as is the case for terrestrial type Ib diamonds. In accordance with other observations, our electron paramagnetic resonance measurements suggest a high H content in presolar diamonds. The presolar diamonds most likely originated in a H‐rich region, an environment in which nanometer‐sized diamonds may be more stable than graphite (Badziag et al., 1990). This adds to the evidence—previously based mainly on the twin microstructures of presolar diamonds (Daulton et al., 1996) and the absence of graphite with the same isotopic composition as presolar diamonds (Anders and Zinner, 1993)—for a homogeneous nucleation of presolar diamonds from a gas phase. Based on our results for detection of diamonds in space, we suggest searching for the N‐induced IR and UV absorption features of type Ib diamonds. Other characteristic diamond features that could also be used to detect diamonds in space are the (‐CH_n) IR absorption features due to H‐coated diamonds, as they are described by Allamandola et al. (1993) and the IR multiphonon absorption features of the diamond lattice. The multiphonon features are very weak (Edwards, 1985), but their intensity increases somewhat with increasing temperature (Collins and Fan, 1954), so perhaps a search for them is not totally hopeless.     

Average temporal signatures of GRBs: applications to theoretical models

New statistical results for cosmic gamma-ray bursts are presented, based on aligned average time profiles. General signatures are discussed of averaged flux/spectrum evolution of the full set of the Second BATSE Catalog GRBs (Fishmanet al., 1995), as well as of the bright and dim subsets. The contradictory conclusions made by Norriset al., (1994) and Mitrofanovet al., (1995a) about the presence or the absence of the time-dilation of dim GRBs are discussed. The well-established effect of hardness/brightness correlation of GRBs is presented. The consequences of these statistical studies are discussed for the cosmological and galactic paradigms of GRBs sources.     

How to characterize terrains on 4 Vesta using Dawn Framing Camera color bands?

We present methods for terrain classification on 4 Vesta using Dawn Framing Camera (FC) color information derived from laboratory spectra of HED meteorites and other Vesta-related assemblages. Color and spectral parameters have been derived using publicly available spectra of these analog materials to identify the best criteria for distinguishing various terrains. We list the relevant parameters for identifying eucrites, diogenites, mesosiderites, pallasites, clinopyroxenes and olivine + orthopyroxene mixtures using Dawn FC color cubes. Pseudo Band I minima derived by fitting a low order polynomial to the color data are found to be useful for extracting the pyroxene chemistry. Our investigation suggests a good correlation (R² = 0.88) between laboratory measured ferrosilite (Fs) pyroxene chemistry vs. those from pseudo Band I minima using equations from Burbine et al. (Burbine, T.H., Buchanan, P.C., Dolkar, T., Binzel, R.P. [2009]. Planetary Science 44, 1331–1341). The pyroxene chemistry information is a complementary terrain classification capability beside the color ratios. We also investigated the effects of exogenous material (i.e., CM2 carbonaceous chondrites) on the spectra of HEDs using laboratory mixtures of these materials. Our results are the basis for an automated software pipeline that will allow us to classify terrains on 4 Vesta efficiently.     

Search for extractable fullerenes in the Allende meteorite

Abstract— In an effort to confirm earlier observations of fullerenes in the Allende meteorite by Becker et al. (1994), we have treated nine separate whole-rock chips of this meteorite with toluene to search for extractable C₆₀ and C₇₀ fullerenes. The analysis was done with high performance liquid chromatography. Less than 1 ppb C₆₀ and C₇₀ were detected in a total of 131.82 g of this meteorite. Becker et al. (1994) have estimated a content of 100 ppm in one Allende sample. Either they have grossly overestimated the C₆₀ content of their sample, or Allende is very inhomogeneous concerning fullerenes.