TDSC-Net:一种基于注意力机制与特征融合的二维恒星光谱分类模型  

TDSC-Net:A Two-Dimensional Stellar Spectra Classification Model Based on Attention Mechanism and Feature Fusion

作　　者：李荣 曹冠龙[1] 蒲源 邱波 王晓敏[1] 闫静 王坤[1] LI Rong;CAO Guan-long;PU Yuan;QIU Bo;WANG Xiao-min;YAN Jing;WANG Kun(Hebei University of Technology,Tianjin 300400,China;Guangdong Baiyun University,Guangzhou 510450,China)

机构地区：[1]河北工业大学,天津300400 [2]广东白云学院,广东广州510450

出　　处：《光谱学与光谱分析》2024年第7期1968-1973,共6页Spectroscopy and Spectral Analysis

基　　金：天津市自然科学基金项目(22JCYBJC00410);国家自然科学基金委员会-中国科学院天文联合基金项目(U1931134)资助。

摘　　要：面对信噪比较低的天体,传统一维光谱的分类效果很差。因此,从二维光谱出发,提出了结合注意力机制的特征融合模型TDSC-Net(two-dimensional spectra classification network)用于恒星分类。TDSC-Net通过完全相同的特征提取层分别对恒星蓝端和红端的二维光谱图像进行特征提取,之后针对这些特征进行融合,然后进行分类。本文实验中的恒星光谱数据选自LAMOST(the large sky area multi-object fiber spectroscopic telescope)数据库,预处理采用Z-score进行光谱归一化,以减少由于光谱流量值差别大造成的模型收敛困难问题。使用精确率(Precision)、召回率(Recall)、F1-score和准确率(Accuracy)四个指标来评估模型性能。实验包括:第一部分利用TDSC-Net进行A、F、G、K、M型恒星分类,以验证利用二维光谱对恒星多分类的可靠性;第二部分将二维光谱按照不同的信噪比区间进行分类,以探究信噪比对分类准确率的影响。第一部分的结果表明,进行五分类总的准确率达到84.3%。其中,A、F、G、K、M各自的分类精度分别为87.0%,84.6%,81.2%,87.4%,89.7%,均优于自行抽谱后的一维光谱分类结果。第二部分的结果表明,即使在SNR<30的低信噪比区间,二维光谱分类准确率仍然可以达到78.9%;而当SNR>30之后,信噪比对光谱分类的影响就不明显了。由此证明了低信噪比时使用二维光谱分类的重要性以及TDSC-Net对恒星光谱分类的有效性。Technology;Guangdong Baiyun University;Traditional one-dimensional spectra perform poorly when classifying celestial objects with a low signal-to-noise ratio(SNR).Therefore,the paper uses two-dimensional spectra and proposes a feature fusion model called TDSC-Net(Two-Dimensional Spectra Classification Network),incorporating an attention mechanism for stellar classification.TDSC-Net employs identical feature extraction layers to get features from the two-dimensional spectra of stars,specifically from the blue and red ends.The extracted features are fused and employed for the classification task.The stellar spectral data in this experiment is selected from the LAMOST(the Large Sky Area Multi-Object Fiber Spectroscopic Telescope)database.They are using Z-score normalization on the spectra to reduce convergence difficulties caused by significant variations in spectral flux values and evaluating the model performance by four metrics:Precision,Recall,F1-score,and Accuracy.The experiments consist of two parts:In the first part,TDSC-Net is employed to classify A,F,G,K,and M-type stars to validate the reliability of using two-dimensional spectra for multi-class stellar classification.In the second part,the two-dimensional spectra are classified based on different SNRs to investigate the impact of SNRs on classification accuracy.The first part's results show that the five-class classification accuracy reaches 84.3%.The classification accuracies of A,F,G,K,and M types are 87.0%,84.6%,81.2%,87.4%,and 89.7%,respectively.These accuracies are higher than the results obtained from one-dimensional spectra classification after spectra extraction.The results of the second part indicate that even in the low SNR(SNR<30),the accuracy of two-dimensional spectra classification can still reach 78.9%.Once the SNR surpasses 30,the impact of SNR on spectra classification becomes less significant.These provide evidence for the importance of using two-dimensional spectra classification in low SNR and demonstrate the effectiveness of TDSC-Net

关 键 词：恒星分类 卷积神经网络 注意力机制 Two-dimensional spectra classification network 

分 类 号：P157.2[天文地球—天文学]