基于FPN-ViT的星系形态分类研究

随着人工智能技术的发展, 利用深度学习方法进行星系形态分类研究取得了较大进展, 但在分类精度、自动化及其星系的空间特征表示上仍然存在不足之处. Vision Transformer (ViT)模型目前在星系形态分类上具有较好的鲁棒性, 但是在处理多尺度图像时存在一定的局限性, 因此提出将特征金字塔(Feature Pyramid Networks, FPN)引入ViT模型(FPN-ViT)中进行星系形态的分类研究中. 结果表明: 基于FPN-ViT模型进行星系形态分类的平均准确率、精确率、召回率以及F1分数等各项评估指标均在95%以上, 与传统的ViT模型相比各项指标均有一定程度的提升. 同时, 在原始星系图像中加入不同程度的高斯噪声和椒盐噪声, 验证FPN-ViT模型对低信噪比数据也能获得较好的分类性能. 此外, 为了对模型进行综合评估, 采用t分布随机邻接嵌入(t-distributed Stohastic Neighbor Embedding, t-SNE)算法对分类结果进行了可视化分析, 能够更加直接地看出FPN-ViT模型对于星系形态分类的效果. 因此, 将FPN网络应用于ViT模型对星系形态的分类研究中是一种全新尝试, 对后续研究具有重要意义.     

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.     

一种低表面亮度星系的自动搜索算法newline--YOLOX-CS

低表面亮度星系(Low Surface Brightness Galaxy, LSBG)的特征对于理解星系整体特征非常重要, 通过现代的机器学习特别是深度学习算法来搜寻扩充低表面亮度星系样本具有重要意义. LSBG因特征不明显而难以用传统方法进行自动和准确辨别, 但深度学习确具有自动找出复杂且有效特征的优势, 针对此问题提出了一种可用于在大样本巡天观测项目中搜寻LSBG的算法---YOLOX-CS (You Only Look Once version X-CS). 首先通过实验对比5种经典目标检测算法并选择较优的YOLOX算法作为基础算法, 然后结合不同注意力机制和不同优化器, 构建了YOLOX-CS的框架结构. 数据集使用的是斯隆数字化巡天(Sloan Digital Sky Survey, SDSS)中的图像, 其标签来自于$\alpha.40$-SDSS DR7 (40%中性氢苜蓿巡天与第7次数据发布的斯隆数字化巡天的交叉覆盖天区)巡天项目中的LSBG, 由于该数据集样本较少, 还采用了深度卷积生成对抗网络(Deep Convolutional Generative Adversarial Networks, DCGAN)模型扩充了实验测试数据. 通过与一系列目标检测算法对比后, YOLOX-CS在扩充前后两个数据集中搜索LSBG的召回率和AP (Average Precision)值都有较好的测试结果, 其在未扩充数据集的测试集中的召回率达到97.75%, AP值达到97.83%, 在DCGAN模型扩充的数据集中, 同样测试集下进行实验的召回率达到99.10%, AP值达到98.94%, 验证了该算法在LSBG搜索中具有优秀的性能. 最后, 将该算法应用到SDSS部分测光数据上, 搜寻得到了765个LSBG候选体.     

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．     

COSMOS场中星系恒星形成的演化研究

基于COSMOS(Cosmic Evolution Survey)/Ultra VISTA(Ultra-deep Visible and Infrared Survey Telescope for Astronomy)场中多波段测光数据,利用质量限选取了红移分布在0z3.5的星系样本.通过UVJ(U-V和V-J)双色图分类判据将星系分类成恒星形成星系(SFGs)和宁静星系(QGs).对于红移分布在0z1.5范围内且M*1011M⊙的QGs来说,该星系在样本中所占比例高于70%.在红移0z3.5范围内,恒星形成星系的恒星形成率(SFR)与恒星质量(M*)之间有着很强的主序(MS)关系.对于某一固定的恒星质量M*来说,星系的SFR和比恒星形成率(s SFR)会随着红移增大而增大,这表明在高红移处恒星形成星系更加活跃,有激烈的恒星形成.相对于低质量的星系来说,高质量的SFGs有较低的s SFR,这意味着低质量星系的增长更多的是通过星系本身的恒星形成.通过结合来自文献中数据点信息,发现更高红移(2z8)星系的s SFR随红移的演化趋势变弱,其演化关系是s SFR∝(1+z)0.94±0.17.     

星系中恒星形成率指针的比较研究

利用赫歇尔空间望远镜的H-ATLAS(Herschel Astrophysical Terahertz Large Area Survey)SDP(Science Demonstration Phase)天区从紫外到亚毫米波段数据,结合星族合成方法和尘埃模型,计算了星系的红外总光度.在此基础上,分别针对强恒星形成星系和弱恒星形成星系,研究了利用紫外光度、红外光度和Hα谱线计算得到的恒星形成率(Star Formation Rate,SFR)的差异以及导致差异的内在物理起因.发现对于恒星形成活动强的星系,这3种恒星形成率指针给出的结果基本一致,弥散较小、只是在高恒星形成率端,利用紫外光度算得的恒星形成率比利用Hα谱线流量算得的恒星形成率略微偏小;而在低恒星形成率端,紫外光度指针偏大于Hα谱线指针;红外光度指针与Hα谱线指针在两端无明显偏差.对弱恒星形成星系,紫外光度、Hα谱线和红外光度3种恒星形成率指针存在明显的差异,且弥散较大.利用紫外光度和Hα谱线计算得到的恒星形成率的弥散和系统偏差随着星系年龄、质量的增加而增大.系统偏差增大的主要原因是利用紫外连续谱斜率β定标恒星形成活动较弱星系的消光时,高估了这些星系的紫外消光,使得消光改正后的紫外光度偏大.另外,MPA/JHU(Max Planck Institute for Astrophysics/Johns Hopkins University)数据库中弱恒星形成星系的恒星形成率SFR(Hα)比真实值偏低.     

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.     

Study of Star/Galaxy Classification Based on the XGBoost Algorithm

Machine learning has achieved great success in many areas today. The lifting algorithm has a strong ability to adapt to various scenarios with a high accuracy, and has played a great role in many fields. But in astronomy, the application of lifting algorithms is still rare. In response to the low classification accuracy of the dark star/galaxy source set in the Sloan Digital Sky Survey (SDSS), a new research result of machine learning, eXtreme Gradient Boosting (XGBoost), has been introduced. The complete photometric data set is obtained from the SDSS-DR7, and divided into a bright source set and a dark source set according to the star magnitude. Firstly, the ten-fold cross-validation method is used for the bright source set and the dark source set respectively, and the XGBoost algorithm is used to establish the star/galaxy classification model. Then, the grid search and other methods are used to adjust the XGBoost parameters. Finally, based on the galaxy classification accuracy and other indicators, the classification results are analyzed, by comparing with the models of function tree (FT), Adaptive boosting (Adaboost), Random Forest (RF), Gradient Boosting Decision Tree (GBDT), Stacked Denoising AutoEncoders (SDAE), and Deep Belief Nets (DBN). The experimental results show that, the XGBoost improves the classification accuracy of galaxies in the dark source classification by nearly 10% as compared to the function tree algorithm, and improves the classification accuracy of sources with the darkest magnitudes in the dark source set by nearly 5% as compared to the function tree algorithm. Compared with other traditional machine learning algorithms and deep neural networks, the XGBoost also has different degrees of improvement.     

Spectral Preprocessing and Its Effect on Galaxy/Quasar Classification

Because of the effects of noise, distortion, observational environment and other factors, some appropriate preprocessing should be made in advance of automatic classification of celestial spectra. We have studied the effect of data format and flux standardization on the automatic classification of sky survey spectra. A basic model adaptable for the order-of-magnitude variation of fluxes is proposed, and the corresponding standardization methods are given. Our experimental results on galaxy and quasar classification show that the logarithmic wavelength data format is better for the automatic spectral classification. By these experiments, the reasonableness of the proposed model and the performances of the given flux standardization methods are verified. Especially, it is noted that the commonly used flux standardization is the worst, among other standardizations, for automatic spectral classification.     

Research on Star/Galaxy Classification Based on Stacking Ensemble Learning

Machine learning has achieved great success in many areas today, but the forecast effect of machine learning often depends on the specific problem. An ensemble learning forecasts results by combining multiple base classifiers. Therefore, its ability to adapt to various scenarios is strong, and the classification accuracy is high. In response to the low classification accuracy of the darkest source magnitude set of stars/galaxies in the Sloan Digital Sky Survey (SDSS), a star/galaxy classification algorithm based on the stacking ensemble learning is proposed in this paper. The complete photometric data set is obtained from the SDSS Data Release (DR) 7, and divided into the bright source magnitude set, dark source magnitude set, and darkest source magnitude set according to the stellar magnitude. Firstly, the 10-fold nested cross-validation method is used for the darkest source magnitude set, then the Support Vector Machine (SVM), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost) algorithms are used to establish the base-classifier model; the Gradient Boosting Decision Tree (GBDT) is used as the meta-classifier model. Finally, based on the classification accuracy of galaxies and other indicators, the classification results are analyzed and compared with the results obtained by the Function Tree (FT), SVM, RF, GBDT, Stacked Denoising Autoencoders (SDAE), Deep Belief Nets (DBN), and Deep Perception Decision Tree (DPDT) models. The experimental results show that the stacking ensemble learning model has improved the classification accuracy of galaxies in the darkest source magnitude set by nearly 10% compared to the function tree algorithm. Compared with other traditional machine learning algorithm, stronger lifting algorithm, and deep learning algorithm, the stacking ensemble learning model also has different degrees of improvement.     

An Automatic Method for Detecting Transients and Variable Sources in AST3 Survey Based on Image Subtraction and Random Forest

AST3-2 (the second Antarctic Survey Telescope) is located in Antarctic Dome A, the loftiest ice dome on the Antarctic Plateau. It produces a huge amount of observational data which require a more efficient data reduction program to be developed. Also the data transmission in Antarctica is much difficult, thus it is necessary to perform data reduction and detect variable and transient sources remotely and automatically in Antarctica, but this attempt is restricted by the unsatisfactory performance of the low power consumption computer in Antarctica. For realizing this purpose, to develop a new method based on the existing image subtraction method and random forest algorithm, taking the AST3-2 2016 dataset as the test sample, becomes an alternative choice. This method performs image subtraction on the dataset, then applies the principle component analysis to extract the features of residual images. Random forest is used as a machine learning classifier, and in the test a recall rate of 97% is resulted for the positive sample. Our work has verified the feasibility and accuracy of this method, and finally found out a batch of candidates for variable stars in the AST3-2 2016 dataset.