Morpho-z: improving photometric redshifts with galaxy morphology

Abstract

We conduct a comprehensive study of the effects of incorporating galaxy morphology information in photometric redshift estimation. Using machine learning methods, we assess the changes in the scatter and outlier fraction of photometric redshifts when galaxy size, ellipticity, Sérsic index and surface brightness are included in training on galaxy samples from the SDSS and the CFHT Stripe-82 Survey (CS82). We show that by adding galaxy morphological parameters to full ugriz photometry, only mild improvements are obtained, while the gains are substantial in cases where fewer passbands are available. For instance, the combination of gr z photometry and morphological parameters almost fully recovers the metrics of 5-band photometric redshifts. We demonstrate that with morphology it is possible to determine useful redshift distribution N(z) of galaxy samples without any colour information. We also find that the inclusion of quasar redshifts and associated object sizes in training improves the quality of photometric redshift catalogues, compensating for the lack of a good star-galaxy separator. We further show that morphological information can mitigate biases and scatter due to bad photometry. As an application, we derive both point estimates and posterior distributions of redshifts for the official CS82 catalogue, training on morphology and SDSS Stripe-82 ugriz bands when available. Our redshifts yield a 68th percentile error of 0.058(1+ z), and a outlier fraction of 5.2 per cent. We further include a deep extension trained on morphology and single i-band CS82 photometry.

Photometric Redshift Estimation with Galaxy Morphology Using Self-organizing Maps

“…Similarly, we conclude that our use of morphology, at its present precision, may not be providing any new information that is not already contained in our 15 bands of photometry. Soo et al (2018) also compare the effects of low-quality versus high-quality morphology by studying galaxy radii measured by the Sloan Digital Sky Survey (SDSS) Stripe-82 survey and by the Canada-France-Hawaii Telescope (CFHT) in Stripe 82 (CS82), the latter of which they assume to be of higher quality due to its 0.6 arcsecond seeing. However, they do not find any improvement in photo-zs when using the CS82 data over the SDSS data.…”

Morpho-photometric redshifts

“…The main idea leading to the network architecture presented in this paper originates from the well-documented astrophysical knowledge that photometric (galaxy-integrated) colors are the primary variables correlating with the spectroscopic redshifts of galaxies, while galaxy morphology is typically of secondary importance (Soo et al 2018). This means that a convnet presented with images (only) will first need to learn to integrate the fluxes of galaxy over their spatial extent to perform well on the redshift regression task before it can focus on secondary morphological details.…”

“…This observation naturally leads to the concept of a mixed MLP-convnet architecture with concatenated features, since the MLP branch should perform well on photometric quantities and the convnet should do well on complementary morphogical information. Ideally, the convnet will learn the best complementary features to the photometric quantities, and in so could outperform the hand-crafted morphological features that have been traditionally used by astronomers (Soo et al 2018, and references therein).…”

“…is the standard deviation of the redshift bias δz and σ68 is the sample-mean 68% percentile spread of the absolute error on the redshift z (see, e.g. Soo et al 2018, for similar implementations). Finally, we report the outlier fraction beyond 3σ68, referred to as f (3σ68).…”

“…In an attempt to improve photo-z estimations, various studies have considered adding morphological information, beyond the usual galaxy-integrated colors, as training input for the machine. 1 Results have been somewhat inconclusive, in the sense that performance gains are not systematic when morphological data is added: performance gains/losses from morphological data depend on details of the task/dataset/method adopted (see Soo et al 2018, and references therein for a detailed account). Reports of gains or losses depending on which combination of morphological features are included are not entirely surprising: feature engineering is a notoriously 'non-linear' process 2 and optimizing over the feature space for best performance is a generally complex enterprise (e.g., .…”