基于DenseNet的天体光谱分类方法  

Classification of Astronomical Spectra Based on DenseNet

作　　者：王奇勋 赵刚[1,2] 范舟 Wang Qixun;Zhao Gang;Fan Zhou(Key Laboratory of Optical Astronomy,Chinese Academy of Sciences,Beijing 100101,China;School of Astronomy and Space Science,University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院光学天文重点实验室,北京100101 [2]中国科学院大学天文与空间科学学院,北京100049

出　　处：《天文研究与技术》2020年第1期85-95,共11页Astronomical Research & Technology

基　　金：国家自然科学基金(11390371)资助.

摘　　要：天体光谱数据的智能处理正由传统机器学习方法逐步转向深度学习,主要采用基于计算机视觉的技术手段。基于计算机视觉领域广泛应用的DenseNet网络结构,针对光谱数据进行修改,建立了适用于光谱数据的一维卷积神经网络模型,解决天体光谱数据分类任务。在验证数据集上,恒星、星系、类星体的F1分数达到了0.9987、0.9127、0.9147,高于传统的神经网络。光谱分类关注区域的可视化结果表明,本文模型可以学习到各类天体对应的特征谱线,具有较强的可解释性。本文方法被用于阿里云天池天文数据挖掘大赛——天体光谱智能分类,并在843支参赛队伍的3次数据评比中获得了2次第一、1次第三的成绩,证明了该模型在保证分类精度的同时,具有极强的鲁棒性、泛化性,适用于光谱的自动分类。Intelligent processing of astronomical spectra data is gradually shifting from traditional machine learning to deep learning,which mainly uses the technology of computer vision.Based on DenseNet network structure,which is widely used in the field of computer vision,a one-dimensional convolution neural network model for spectral data is established to solve the classification task of celestial spectral data.The F1 scores of stars,galaxies and quasars are 0.9987,0.9127 and 0.9147 respectively in the validated data set.The visualization results of the regions of interest in spectral classification show that the proposed model can learn the characteristic spectral lines of celestial bodies which has interpretability.This method is applied to the intelligent classification of celestial spectrum in Alibaba Cloud Tianchi Astronomical Data Mining Competition.We won the first prize two times and the third prize one time in 843 teams in three data evaluations,and the result proves that the model has robustness and generalization,and is suitable for automatic classification of spectra.

关 键 词：卷积神经网络 光谱分类 数据挖掘 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]