Classification of Astronomical Spectra Based on DenseNet

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.