Finding Strong Gravitational Lenses in the DESI DECam Legacy Survey

Abstract

We perform a semi-automated search for strong gravitational lensing systems in the 9,000 deg 2 Dark Energy Camera Legacy Survey (DECaLS), part of the DESI Legacy Imaging Surveys (Dey et al.). The combination of the depth and breadth of these surveys are unparalleled at this time, making them particularly suitable for discovering new strong gravitational lensing systems. We adopt the deep residual neural network architecture (He et al.) developed by Lanusse et al. for the purpose of finding strong lenses in photometric surveys. We compile a training set that consists of known lensing systems in the Legacy Surveys and DES as well as non-lenses in the footprint of DECaLS. In this paper we show the results of applying our trained neural network to the cutout images centered on galaxies typed as ellipticals (Lang et al.) in DECaLS. The images that receive the highest scores (probabilities) are visually inspected and ranked. Here we present 335 candidate strong lensing systems, identified for the first time.

Galaxy and Mass Assembly: A Comparison between Galaxy–Galaxy Lens Searches in KiDS/GAMA

“…Similar efforts, mostly machine learning, have found strong lens candidates in deep imaging surveys (e.g., Jacobs et al 2019, Speagle et al 2019, and Huang et al 2020b. Figure 10 shows the GAMA equatorial lens candidates of this work and includes two candidates previously identified by DECaLS, SLACS (Bolton et al 2008a) in the GAMA equatorial regions that had a match in R.A./decl.…”

“…Machine learning is gaining popularity as a method for identifying galaxy-galaxy lens candidates, e.g., in Subaru Hyper-Supreme Cam (Speagle et al 2019), DECAM (Huang et al 2020b), and Dark Energy Survey data (Jacobs et al 2019). Petrillo et al (2017Petrillo et al ( , 2019aPetrillo et al ( , 2019b introduced and developed a machine-learning technique to visually identify strong lens candidates by training the convolutional neural networks to recognize the characteristic arcs that appear next to a lensing elliptical galaxy using simulated images as their training set.…”

“…Each method's catalog occupies a unique range of both characteristics. Left: the two candidates that overlap between GalaxyZoo citizen-science and LinKS machine-learning samples are marked with square and diamond outlines, and we point out candidates G93310 and G136604 that were previously identified in DECaLS (Huang et al 2020b) and SLACS (Bolton et al 2008a), respectively. Right: error bars represent uncertainty in stellar mass estimates.…”

“…The left figure shows SLACS (Bolton et al 2006), S4TM (Shu et al 2017) from SDSS-DR7 (Abazajian et al 2009), and BELLS (Brownstein et al 2012) spectroscopic searches compared to the GAMA spectroscopy candidates selected in this work from Holwerda et al (2015). The right figure shows the LinKS survey from Petrillo et al (2019a) and the candidates presented in this paper from GAMA/KiDS GalaxyZoo citizen science, compared with deep imaging searches in DES (Jacobs et al 2019) and DECaLS (Huang et al 2020b). We note that SDSS spectroscopic searches aimed for clean rather than complete samples (requiring high signal-to-noise ratios) and GAMA emphasizes completeness throughout the survey.…”

“…With the Galaxy and Mass Assembly survey (GAMA) spectroscopy-selected sample (Holwerda et al 2015), a greater range in lens masses is now available. However, the drive for much larger samples has led to increased searches using machine learning to constrain ΛCDM in detail (Petrillo et al 2017(Petrillo et al , 2019a(Petrillo et al , 2019bJacobs et al 2019;Speagle et al 2019;Huang et al 2020aHuang et al , 2020bLi et al 2020).…”