Classification of Astrophysical Spectra using Deep Learning Pre-Trained Algorithms

Abstract

The study of spectroscopy allows us to characterize the properties of astrophysical objects, such as stars, quasars, galaxies, and the intergalactic medium. Currently, a collection of space and ground-based telescopes map the sky every night, increasing the volume of data exponentially over time. This increment in the datasets pushes astronomers to classify the spectra of observed objects efficiently. In this work we implement the pre-trained algorithms: Inception V3, ResNet 50 and MNIST, based on Computer Vision (CV) and Convolutional Neural Networks (CNN). We compare the classification performance of spectra taken from the Sloan Digital Sky Survey (SDSS) in its Data Release DR12, with the Baryon Oscillation Spectroscopic Survey (BOSS) spectrograph with a sample of 300000 spectra. In the learning process, the hyperparameters associated with the model input such as kernel, stride, epochs, class and initial layer were adjusted, moreover each model was evaluated by implementing the metrics of accuracy, precision, and sensitivity. The classification results show that the best classifier for quasar spectra is ResNet 50 with a performance of more than 60\% compared to its yield with galaxies and stars. In addition, a low loss rate was obtained in the case of star classification with the Inception V3 model. This study confirms that algorithms based on CV and CNNs are very powerful for classifying astrophysical spectra.