Galaxy Light Profile Convolutional Neural Networks (GaLNets). I. Fast and Accurate Structural Parameters for Billion-galaxy Samples

Abstract

Next-generation large sky surveys will observe up to billions of galaxies for which basic structural parameters are needed to study their evolution. This is a challenging task that, for ground-based observations, is complicated by seeing-limited point-spread functions (PSFs). To perform a fast and accurate analysis of galaxy surface brightness, we have developed a family of supervised convolutional neural networks (CNNs) to derive Sérsic profile parameters of galaxies. This work presents the first two Galaxy Light profile CNNs (GaLNets) of this family. The first one is trained using galaxy images only (GaLNet-1), and the second is trained with both galaxy images and the local PSF (GaLNet-2). We have compared the results from GaLNets with structural parameters (total magnitude, effective radius, Sérsic index, etc.) derived from a set of galaxies from the Kilo-Degree Survey by 2DPHOT as a representative of the “standard” PSF-convolved Sérsic fitting tools. The comparison shows that GaLNet-2 can reach an accuracy as high as that of 2DPHOT, while GaLNet-1 performs worse because it misses the information from the local PSF. Both GaLNets are three orders of magnitude faster than standard methods in terms of computational speed. This first application of CNNs to ground-based galaxy surface photometry shows that they are promising tools to perform parametric analyses of very large galaxy samples, like the ones expected from the Vera Rubin/LSST surveys. However, GaLNets can be easily modified for space observations from Euclid and the China Space Station Telescope.

MORPHOFIT: An automated galaxy structural parameters fitting package

“…This rapidly growing sample of galaxies with quality imaging has sparkled renovated interests in the development of codes for the measurement of galaxy structural parameters and galaxy morphological classification that are accurate, time efficient, and require very few input from the user. These codes involve the use of non-parametric fitting (CAS Conselice, 2003, GINI Lotz et al, 2004, MORFOMETRYKA Ferrari et al, 2015), parametric fitting (GIM2D Simard et al, 2002, BUDDA de Souza et al, 2004, GASP2D Méndez-Abreu et al, 2008, PYMORPH Vikram et al, 2010, GALAPAGOS Häußler et al, 2011Häußler et al, 2013, IMFIT Erwin, 2015, GALIGHT Ding et al, 2021, galapagos-2 Häußler et al, 2022 or machine learning methodologies ( DEEPLEGATO Tuccillo et al, 2018, GAMORNET Ghosh et al, 2020, GALNETS Li et al, 2022).…”

The Dawes Review 10: The impact of deep learning for the analysis of galaxy surveys

“…Top row shows the results as a function of magnitude while the bottom row is as a function of radius. Figure adapted from Li et al (2021). …”

“…More recently, Li et al (2021) attempted a similar approach applied to ground-based observations. The training is still done on simulations but with realistic backgrounds.…”

Holismokes

“…As this leads to no improvement, we considered networks accepting eight frames instead of four, and added the PSF images directly as input images. To this end, we subsampled the PSF with a linear interpolation to 64 × 64 pixels, matching the size of the images, and passed them independently of the images through a mirrored branch of convolutional layers, combining the intermediate outputs directly before the FC layers as suggested by Maresca et al (2021) and Li et al (2022). Again, no improvement is seen with this option, possibly because such networks, with their relatively small kernel sizes, perform well in pattern recognition but perhaps not as well in analyzing very similar and completely smooth images like a PSF.…”