Machine learning technique for morphological classification of galaxies from the SDSS. III. The CNN image-based inference of detailed features

Abstract

This paper follows a series of our works on the applicability of various machine learning methods to morphological galaxy classification (Vavilova et al., 2021, 2022). We exploited the sample of ~315800 low-redshift SDSS DR9 galaxies with absolute stellar magnitudes of −24m < Mr < −19.4m at 0.003 < z < 0.1 redshifts as a target data set for the CNN classifier. Because it is tightly overlapped with the Galaxy Zoo 2 (GZ2) sample, we use these annotated data as the training data set to classify galaxies into 34 detailed features. In the presence of a pronounced difference in visual parameters between galaxies from the GZ2 training data set and galaxies without known morphological parameters, we applied novel procedures, which allowed us for the first time to get rid of this difference for smaller and fainter SDSS galaxies with mr < 17.7. We describe in detail the adversarial validation technique as well as how we managed the optimal train-test split of galaxies from the training data set to verify our CNN model based on the DenseNet-201 realistically. We have also found optimal galaxy image transformations, which help increase the classifier’s generalization ability. We demonstrate for the first time that implication of the CNN model with a train-test split of data sets and size-changing function simulating a decrease in magnitude and size (data augmentation) significantly improves the classification of smaller and fainter SDSS galaxies. It can be considered as another way to improve the human bias for those galaxy images that had a poor vote classification in the GZ project. Such an approach, like autoimmunization, when the CNN classifier, trained on very good galaxy images, is able to retrain bad images from the same homogeneous sample, can be considered co-planar to other methods of combating such a human bias. The most promising result is related to the CNN prediction probability in the classification of detailed features. The accuracy of the CNN classifier is in the range of 83.3—99.4 % depending on 32 features (exception is for “disturbed” (68.55 %) and “arms winding medium” (77.39 %) features). As a result, for the first time, we assigned the detailed morphological classification for more than 140000 low-redshift galaxies, especially at the fainter end. A visual inspection of the samples of galaxies with certain morphological features allowed us to reveal typical problem points of galaxy image classification by shape and features from the astronomical point of view. The morphological catalogs of low-redshift SDSS galaxies with the most interesting features are available through the UkrVO website (http://ukr-vo.org/galaxies/) and VizieR.