DEEP LEARNING FOR MORPHOLOGICAL CLASSIFICATION OF GALAXIES FROM SDSS

Abstract

We present the results of applying deep convolutional neural network to the images of redshift-limited ( z < 0 . 1) sample of ∼ 300000 galaxies from the SDSS DR9. We aimed to classify galaxies into the two classes: Elliptical and Spiral. To create the training sample, we used a set of ∼ 6000 galaxies from our previous work with visually inspected morphologi- cal types, and also added 80000 well-confirmed galaxies from Galaxy Zoo 2 dataset, that were also classified visually. With a given sample of ∼ 86000 galaxies, we used the deep neural network, namely Xception, to provide a classification of g-r-i composite images (25 arcsec in each axis in size) of galaxies. Keeping in the mind a relatively small training dataset, we provided the data augmentation (horizontal and vertical flips, random shifts on ± 10 pixels , and rotations within 180 degrees), that was randomly applied to the images during learning. The data augmentation is a key technique within our algorithm to display the variative nature of the observed galaxies, and avoid overfitting problem. We compared our classification result with the Support Vector Machine (SVM) classification performed on the SDSS photometric data (absolute magnitudes, colour indices, inverse concentration index, ratios of semiaxes, etc.), and proposed a method to learn the benefits from both approaches (Deep Learning and photometric classification). We show the common mistakes of both algorithms, and propose to stack these two approaches to block these mistakes, with a main goal to increase the overall classification quality of SDSS galaxies.