Applying Deep Neural Networks (DNN) for Measuring Photometric Redshifts from Galaxy Images: Preliminary Study

In the cosmological and extragalactic study, distance to a galaxy is an important parameter, by knowing the distance, we can find any other physical parameter such as mass, luminosity, star formation rate, and metallicity. By applying the specific cosmological model, we can measure a distance from the redshift. The exact redshift can only measure by using the spectroscopic technique (Doppler effect), but spectroscopic observation limited to brighter objects and numbers of objects in a single field of view (FoV). While photometric observation can capture fainter objects and more objects in a single FoV. Measurements of photometric redshift could have done by comparing the SED curves of the elliptical galaxy with known spectroscopic redshifts from other elliptical galaxies which we want to find the photometric redshift. Another method is to do linear or non-linear regression, by assuming the redshift is a function of magnitude in each band-pass filter. Therefore, we propose a technique that using full galaxy images in each measured bands and machine learning method for measuring photometric redshift. We pass entire multi-band galaxy images into the machine learning architecture to get an estimated redshift. In this work, we use galaxies images at 0 <= z <= 1 from SDSS DR 10 as the datasets and we use DenseNet, one of the Deep Neural Networks (DNN) architecture.