Survey of Gravitationally-lensed Objects in HSC Imaging (SuGOHI)

Abstract

Context. Strong lenses are extremely useful probes of the distribution of matter on galaxy and cluster scales at cosmological distances, however, they are rare and difficult to find. The number of currently known lenses is on the order of 1000. Aims. The aim of this study is to use crowdsourcing to carry out a lens search targeting massive galaxies selected from over 442 square degrees of photometric data from the Hyper Suprime-Cam (HSC) survey. Methods. Based on the S16A internal data release of the HSC survey, we chose a sample of ∼300 000 galaxies with photometric redshifts in the range of 0.2 < zphot < 1.2 and photometrically inferred stellar masses of log M* > 11.2. We crowdsourced lens finding on this sample of galaxies on the Zooniverse platform as part of the Space Warps project. The sample was complemented by a large set of simulated lenses and visually selected non-lenses for training purposes. Nearly 6000 citizen volunteers participated in the experiment. In parallel, we used YATTALENS, an automated lens-finding algorithm, to look for lenses in the same sample of galaxies. Results. Based on a statistical analysis of classification data from the volunteers, we selected a sample of the most promising ∼1500 candidates, which we then visually inspected: half of them turned out to be possible (grade C) lenses or better. By including lenses found by YATTALENS or serendipitously noticed in the discussion section of the Space Warps website, we were able to find 14 definite lenses (grade A), 129 probable lenses (grade B), and 581 possible lenses. YATTALENS found half the number of lenses that were discovered via crowdsourcing. Conclusions. Crowdsourcing is able to produce samples of lens candidates with high completeness, when multiple images are clearly detected, and with higher purity compared to the currently available automated algorithms. A hybrid approach, in which the visual inspection of samples of lens candidates pre-selected by discovery algorithms or coupled to machine learning is crowdsourced, will be a viable option for lens finding in the 2020s, with forthcoming wide-area surveys such as LSST, Euclid, and WFIRST.

Holismokes

“…The PS1 lens candidates were cross-matched with galaxyscale strong lenses from previous searches with the status of spectroscopally confirmed or candidate systems, by building upon and expanding the current MasterLens database. We compiled grade A and B systems (or equivalent) from a number of recent searches based on optical and near-infrared imaging from DES (Diehl et al 2017;Jacobs et al 2019b,a), CFHTLS (More et al 2016;Jacobs et al 2017), KiDS (Petrillo et al 2017(Petrillo et al , 2019Li et al 2020), DESI (Huang et al 2020a), and HSC (Wong et al 2018;Sonnenfeld et al 2018Sonnenfeld et al , 2020Jaelani et al 2020). Since our network may also be sensitive to lensed quasars with colors and configurations similar to our mock lenses, we also cross-matched with the all-sky database of ∼220 confirmed lensed quasars in the literature (Lemon et al 2019, and references therein) 12 , and with previously identified lensed quasars from HSC (Chan et al 2020) and from PS1 (Rusu et al 2019).…”

HOLISMOKES. VIII. High-redshift, strong-lens search in the Hyper Suprime-Cam Subaru Strategic Program

“…On the other hand, Jacobs et al (2019b) simulated 10,000 z > 0.8 elliptical galaxies with lensed features superimposed, and found that 90% of the simulated lenses can be recovered with these two colour-colour cuts. In addition, we examine the colour distributions of strong-lens candidates discovered in the HSC footprint by the Survey of Gravitationally-lensed Objects in HSC Imaging (SuGOHI) project (Sonnenfeld et al 2018;Wong et al 2018;Chan et al 2020;Jaelani et al 2020;Sonnenfeld et al 2020). Every SuGOHI strong-lens candidate is assigned a grade from A (definite), B (probable), or C (possible) and a lens type from GG (galaxy-galaxy), GQ (galaxy-quasar), CG (cluster/group-galaxy), or CQ (cluster/group-quasar).…”

“…They span a wide redshift range from 0.2 to 1.0 2 . We note that candidates from Sonnenfeld et al (2020) are not considered here because some GG strong-lens candidates therein are actually cluster-or group-scale lenses. Among the selected 99 SuGOHI strong-lens candidates, 92 (or ≈ 93%) further pass the colourcolour cuts in criteria 13-14.…”

“…To construct the non-lens examples for training and validation, we first randomly select 48,213 objects from the parent sample. To further clean this subset, we crossmatch them with a sample of 10,241 known strong lenses and strong-lens candidates (referred to as the known strong lens compilation hereafter) compiled from the literature (e.g., Diehl et al 2017;Sonnenfeld et al 2018;Wong et al 2018;Petrillo et al 2019;Jacobs et al 2019b,a;Chan et al 2020;Jaelani et al 2020;Sonnenfeld et al 2020;Huang et al 2020Huang et al , 2021Cañameras et al 2020;Li et al 2020;Cañameras et al 2021;Li et al 2021;Rojas et al 2021;Savary et al 2021) using a matching radius of 30 arcsecs and remove the 114 matches. Considering the typical lensing rate of 10 −4 -10 −3 (e.g., Browne et al 2003;Bolton et al 2004;Oguri & Marshall 2010;Treu 2010), the remaining 48,099 objects are expected to be sufficiently pure.…”

Holismokes

“…We develop here new supervised neural networks for automated selection of galaxy-scale strong lenses in large-scale multiband surveys, using the Hyper-Suprime Cam Subaru Strategic Program (HSC-SSP; Aihara et al 2018) for testing on images approaching the expected depth and quality of the Rubin Observatory Legacy Survey of Space and Time (LSST) final stacks (see Ivezić et al 2019). Previous non-machine learning identification of galaxy-, group-, and cluster-scale lenses from the Survey of Gravitationally lensed Objects in HSC Imaging (SuGOHI; Sonnenfeld et al 2018Sonnenfeld et al , 2019Sonnenfeld et al , 2020Wong et al 2018;Chan et al 2020;Jaelani et al 2020Jaelani et al , 2021, hereafter S18, S20, W18 for SuGOHI-g) offers an independent observational set to test our classification completeness. In this Letter we validate our deep learning pipeline and present new high-confidence galaxy-scale lens candidates with wide image separations from the Wide layer of the HSC survey, which is only 1 mag shallower than LSST ten-year stacks, and has sufficient sky coverage for lens searches.…”