LAMOST的“Unknown”光谱分类研究:ODS-YOLOv7模型  

The“Unknown”Spectral Classification Study of LAMOST:ODS-YOLOv7 Model

作　　者：王晓敏[1] 高军萍[1] 蒲源 邱波 张健楠[3] 闫静 李荣 WANG Xiao-min;GAO Jun-ping;PU Yuan;QIU Bo;ZHANG Jian-nan;YAN Jing;LI Rong(Hebei University of Technology,Tianjin 300400,China;Guangdong Baiyun University,Guangzhou 510450,China;National Astronomical Observatories,Chinese Academy of Sciences,Beijing 100012,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 [3]中国科学院国家天文台,北京100012

出　　处：《光谱学与光谱分析》2024年第7期1960-1967,共8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：天体识别是天文新发现和深入研究天体的基础。在LAMOST DR8 v1.0发布的低分辨率光谱数据中有约53万条因没有类别标签而被命名为“Unknown”的光谱,其中有88.56%的光谱信噪比在0~10之间,对这批光谱的研究分析将增加LAMOST的有效数据产出。该研究为“Unknown”光谱的分类设计了一种ODS-YOLOv7模型。它是一种端到端的类别预测模型,通过添加一维卷积注意力模块以提高光谱识别能力。在经过一批信噪比在0~10之间的已知类别光谱训练后,ODS-YOLOv7可以学习到低信噪比光谱的有效特征,进而实现对“Unknown”光谱的类别预测。实验表明,该模型在已知类别标记的低信噪比恒星、星系、类星体的光谱识别中,F1-score分别为0.98、0.95、0.95;同时在与传统算法KNN、RF、DT、SVM和深度学习算法1D CNN、1DSSCNN、ResNet、DenseNet、VIT对比实验中取得相对最好的效果。实验结果还给出了ODS-YOLOv7模型对DR8 v1.0中信噪比在0~10的“Unknown”光谱预测置信度分布,在预测类别为恒星、星系、类星体任务中,有92%的分类置信度在60%以上。为保证模型输出质量,本文只选取分类置信度大于99%的光谱类别作为输出结果。以此为依据,在DR8 v1.0和DR9 v0发布的全部“Unknown”光谱中分别有37.19%和47.03%被ODS-YOLOv7模型预测出类别。此外,还增加人工认证以检验该模型预测的正确性。为提升模型的可解释性,参照了二维图像特征可视化的Grad-CAM方法,将其改进为适合于可视化一维光谱数据特征的算法。其结果表明该模型可自动关注到不同的分类特征,使得该模型非常适用于低信噪比“Unknown”光谱的类别预测。Identifying celestial spectra is essential for making new astronomical discoveries and conducting detailed studies of celestial objects.The LAMOST DR8 v1.0 release of low-resolution spectral data contains approximately 530000 spectra named“Unknown”.The reason is that they have no category labels.And 88.56% of these spectra have signal-to-noise ratios between 0 and 10.Therefore,the effective output of LAMOST will increase if we analyze these spectra.In this paper,we propose an ODS-YOLOv7 model to deal with the problem of the“Unknown”spectral classification.It is an end-to-end category prediction model and is suitable for one-dimensional spectra.We also add a one-dimensional convolutional attention module to improve the accuracy of spectra recognition.After training on a set of known category spectra with signal-to-noise ratios between 0 and 10,the ODS-YOLOv7 model can learn the effective features of the low signal-to-noise spectra.Thus,it can enable us to predict“Unknown”spectra.Experiments show that the model has an F1-score of 0.98,0.95,and 0.95 for the spectral identification of low signal-to-noise stars,galaxies,and quasars spectra with known labels.In the meantime,ODS-YOLOv7 obtains the best results in comparison experiments with traditional algorithms KNN,RF,DT,SVM,and deep learning algorithms 1D CNN,1DSSCNN,ResNet,DenseNet,and VIT.The experimental results also give confidence in the predictions of the ODS-YOLOv7 model for the“Unknown”spectra in DR8 v1.0,with 92% of the confidence levels above 60%.To ensure the quality of the model output,only spectral categories with a prediction confidence level greater than 99% are selected as output in this paper.Ultimately,37.19% and 47.03% of the“Unknown”spectra released in DR8 v1.0 and DR9 v0,respectively,are predicted by this model.In addition,the paper tests the accuracy of the model's predictions using manual authentication.To improve the interpretability of the model,the paper takes the Grad-CAM method for two-dimensional image visualisation.It

关 键 词：Unknown光谱分类 ODS-YOLOv7 低信噪比 LAMOST 特征可视化 

分 类 号：P157.2[天文地球—天文学]