HOLISMOKES -- II. Identifying galaxy-scale strong gravitational lenses in Pan-STARRS using convolutional neural networks
R. Canameras, S. Schuldt, S. H. Suyu, S. Taubenberger, T. Meinhardt, L. Leal-Taixe, C. Lemon, K. Rojas, E. Savary
AAstronomy & Astrophysics manuscript no. lens˙search˙ps1˙arxiv c (cid:13)
ESO 2020April 29, 2020
HOLISMOKES – II. Identifying galaxy-scale strong gravitationallenses in Pan-STARRS using convolutional neural networks
R. Ca˜nameras , S. Schuldt , , S. H. Suyu , , , S. Taubenberger , T. Meinhardt , L. Leal-Taix´e , C. Lemon , K. Rojas ,E. Savary (A ffi liations can be found after the references) Received / Accepted
ABSTRACT
We present a systematic search for wide-separation (with Einstein radius θ E (cid:38) . (cid:48)(cid:48) ), galaxy-scale strong lenses in the 30 000 deg of the Pan-STARRS 3 π survey on the Northern sky. With long time delays of a few days to weeks, such systems are particularly well suited for catchingstrongly lensed supernovae with spatially-resolved multiple images and open new perspectives on early-phase supernova spectroscopy and cos-mography. We produce a set of realistic simulations by painting lensed COSMOS sources on Pan-STARRS image cutouts of lens luminous redgalaxies (LRGs) with known redshift and velocity dispersion from SDSS. First of all, we compute the photometry of mock lenses in gri bandsand apply a simple catalog-level neural network to identify a sample of 1 050 207 galaxies with similar colors and magnitudes as the mocks.Secondly, we train a convolutional neural network (CNN) on Pan-STARRS gri image cutouts to classify this sample and obtain sets of 105 760and 12 382 lens candidates with scores p CNN > . > .
9, respectively. Extensive tests show that CNN performances rely heavily on the designof lens simulations and choice of negative examples for training, but little on the network architecture. The CNN correctly classifies 14 out of 16test lenses, which are previously confirmed lens systems above the detection limit of Pan-STARRS. Finally, we visually inspect all galaxies with p CNN > . z ∼ . z s = . z d = . / or spectro-scopic follow-up will be required to validate Pan-STARRS lens candidates and derive strong lensing models. We also expect that the e ffi cient andautomated two-step classification method presented in this paper will be applicable to the ∼ gri stacks from the Rubin ObservatoryLegacy Survey of Space and Time (LSST) with minor adjustments. Key words. gravitational lensing: strong – data analysis: methods
1. Introduction
Strongly lensed systems with time-variable sources providecompetitive probes of the Hubble constant H that are inde-pendent from CMB observations (Planck Collaboration 2018)and the local distance ladder (Riess et al. 2019; Freedman et al.2019, 2020), and allow one to assess the significance of the cur-rent H tension. The H0LiCOW and COSMOGRAIL projects(e.g., Suyu et al. 2017; Courbin et al. 2018) have recently es-tablished the capacity of combining time-delay measurementsand robust strong lensing models to constrain H , and mea-sured H = . + . − . km s − Mpc − in flat Λ CDM cosmologyusing six lensed quasars (Wong et al. 2019). The seventh lenshas been analysed by the STRIDES collaboration (Shajib et al.2020; Buckley-Geer et al. 2020), and a detailed study of sys-tematic e ff ects is presented by Millon et al. (2019) as part ofthe TDCOSMO organisation. Moreover, the first two stronglylensed supernovae (SNe) with spatially-resolved multiple im-ages have been detected in recent years, one core-collapse SNbehind the strong lensing cluster MACS J1149.5 + H measure-ments with lensed SNe. Such systems are indeed well suitedfor time-delay measurements given the smooth, non-erratic SNelight curves which require shorter high-cadence monitoring thanlensed quasars, and the possibility to reduce microlensing ef- fects by focusing on the early, achromatic expansion phase afew weeks after explosion, and by using color light curves (Suyuet al. 2020; Huber et al. 2019; Bonvin et al. 2019; Goldstein et al.2018, Huber et al., 2020, in prep.;). For lensed type Ia SNe, thestandardizable intrinsic peak luminosity of the source is also avaluable input for breaking the mass-sheet degeneracy in lensmass models (Falco et al. 1985). Constraints on H have alreadybeen derived with SN Refsdal (Grillo et al. 2018, 2020) and il-lustrate the great potential of such measurements, in particularfor galaxy-scale strong lens systems that have simpler lens massdistributions than galaxy clusters. Lensed SNe with adequate im-age separations providing time delays of a few days to weeksare particularly promising and are relatively less sensitive to mi-crolensing e ff ects (Suyu et al. 2020; Huber et al. 2019).Besides precise measurements of the Hubble constant,strongly lensed SNe open an interesting window on early-phaseSN studies. Multiply-imaged SNe detected from the first imagecan be combined with strong lensing models to predict the timedelays and future SN reappearance (as was done for SN Refsdal)in order to trigger follow-up observations within a few days ofexplosion, which is currently not feasible for unlensed SNe be-yond the local universe due to their late discovery near peak lu-minosity. Such early-phase studies are particularly valuable totackle the progenitor problem of type Ia SNe and disentanglethe single-degenerate (Whelan & Iben 1973), double-degenerate(Tutukov & Yungelson 1981), and additional scenarios that have a r X i v : . [ a s t r o - ph . GA ] A p r . Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search been extensively debated over the last decades. For core-collapseSNe, these observations are important to characterize the pro-genitor properties and compare with current stellar evolutionmodels. Early-phase spectroscopy of type II SNe would yieldnovel constraints on the mass-loss history just before explosion.We recently initiated the Highly Optimised LensingInvestigations of Supernovae, Microlensing Objects, andKinematics of Ellipticals and Spirals (HOLISMOKES, Suyuet al. 2020) programme to address these fundamental questionson stellar physics and cosmology. The number of strongly lensedSNe is expected to grow over the next few years, thanks to theon-going Zwicky Transient Facility (ZTF, Masci et al. 2019)high-cadence survey on the northern hemisphere and the forth-coming Rubin Observatory Legacy Survey of Space and Time(LSST, Ivezi´c et al. 2019) on the south. Oguri & Marshall (2010)predict 45 strongly lensed type Ia SNe over the 10 years of LSST,which corresponds to a few events for the shallower ZTF survey.This assumes a selection from spatially-resolved multiple im-ages targeting the most useful wide-separation systems. Usingcomplementary selection techniques solely based on magnifica-tion of SN Ia light curves, Goldstein & Nugent (2017) predict10 to 20 times more lensed SN Ia albeit mostly with small im-age separations (see also Wojtak et al. 2019). Importantly, newlensed SNe candidates have to be selected early enough to startthe follow-up sequence in a timely manner. One way is to extendthe numerous, successful searches of galaxy-scale strong lensesthat were traditionally conducted on surveys with optimal imagequality (e.g. More et al. 2016; Sonnenfeld et al. 2018), to surveyswith largest sky coverage, in order to quickly identify transientsmatching the position of background lensed sources. Ultimately,these lens finding pipelines will be directly applicable to the deepLSST stacks which are expected to yield approximately a hun-dred thousand new systems (Collett 2015).Galaxy-scale strong gravitational lenses without time-variable sources also provide valuable insights into the lens to-tal mass distributions, including the inner dark-matter fractions(e.g., Gavazzi et al. 2007; Grillo et al. 2009; Sonnenfeld et al.2015; Schuldt et al. 2019), the slopes of the total and dark-mattermass density profiles (e.g., Treu & Koopmans 2002; Koopmanset al. 2009; Barnab`e et al. 2011; Shu et al. 2015), and the spa-tial extent of dark-matter halos (e.g., Halkola et al. 2007; Suyu &Halkola 2010). Such systems play a crucial role in characterizingthe lens stellar initial mass function (IMF), a major ingredient forstellar mass estimates, as a function of galaxy physical properties(e.g., Ca˜nameras et al. 2017b; Barnab`e et al. 2013; Sonnenfeldet al. 2019), and they are well-suited to search for dark-mattersubstructures (e.g., Vegetti et al. 2012; Hezaveh et al. 2016;Ritondale et al. 2019). Moreover, high magnification factors pro-vide unique diagnostics on the local interstellar medium physicalconditions in lensed high-redshift galaxies and on the local feed-back mechanisms driving their evolution (e.g., Danielson et al.2011; Ca˜nameras et al. 2017a; Cava et al. 2018).Strong lensing events are rare, about 1 in 1000 for high-resolution space-based imaging (e.g., Marshall et al. 2009)and down to about 1 in 10 for seeing-limited ground-baseddata (e.g., Jacobs et al. 2019a), and their identification thusrequires dedicated and automated methods. For instance, arc-finder algorithms (e.g., Gavazzi et al. 2014; Sonnenfeld et al.2018) and citizen-science classification projects (S pace W arps ,Marshall et al. 2016) have been developed over the last decade.In particular, convolutional neural networks (CNNs) are super-vised machine-learning algorithms optimized to image analysis(LeCun et al. 1998) that have proven to outperform other clas-sification techniques and that require little pre-processing. They are very e ffi cient to peer into large imaging data sets and havebeen increasingly used in the field of astronomy over the lastfive years. These studies have established the ability of CNNsin recognizing galaxy morphologies (Dieleman et al. 2015), in-cluding the key features of strong gravitational lenses (Metcalfet al. 2019). Several CNN searches for new strong lens can-didates have focused on ground-based imaging data, from theCFHTLS (Jacobs et al. 2017), KiDS DR3 (Petrillo et al. 2017)and DR4 (Petrillo et al. 2019; Li et al. 2020), DES Year 3 (Jacobset al. 2019b,a), or the DESI DECam Legacy survey (Huanget al. 2019). E ffi cient classification pipelines using deep neu-ral networks have also been developed and tested on simulatedEuclid and LSST images to prepare for these forthcoming sur-veys which will tremendously increase the number of detectablestrong lensing systems (Lanusse et al. 2018; Schaefer et al. 2018;Davies et al. 2019; Cheng et al. 2019; Avestruz et al. 2019).In the meantime, no systematic searches of galaxy-galaxystrong lenses have so far taken advantage of the Pan-STARRSimaging covering the entire Northern sky. With this survey,Berghea et al. (2017) identified a strongly lensed QSO forminga quadruple system by cross-matching the position of a variableAGN selected in the mid-infrared. More recently, Rusu et al.(2019) performed a more systematic search of lensed QSOsby applying color / magnitude cuts and visually inspecting Pan-STARRS image cutouts of AGN candidates from the Wide-fieldInfrared Survey Explorer (Secrest et al. 2015). In this paper, weperform a comprehensive search for galaxy-scale strong lensingsystems with luminous red galaxies (LRGs) as deflectors andtypical high-redshift galaxies as background sources, using theextended footprint of nearly 30 000 deg of the Pan-STARRS3 π survey on the Northern sky. The automated pipeline, basedon a catalog-level pre-selection of galaxies and a convolutionalneural network, results in a ranked list of candidates which wefurther inspect visually to select those with higher confidence forspectroscopic follow-up.The outline of the paper is as follows. In Sect. 2 and 3, wegive a short overview of the Pan-STARRS surveys and the over-all search methodology and, in Sect. 4, we describe the sim-ulation of strong lenses. In Sect. 5, we present the networksand training processes, and we extensively test the CNN per-formance. In Sect. 6 we finally apply the CNN to pre-selectedPan-STARRS image cutouts, provide the list of strong lens can-didates from visual inspection, and characterize their overallproperties. Our main conclusions appear in Sect. 7. Throughoutthis work, we adopt the flat concordant Λ CDM cosmology with Ω M = . Ω Λ = − Ω M (Planck Collaboration XIII 2016),and with H =
72 km s − Mpc − (Bonvin et al. 2017).
2. The Pan-STARRS1 survey
The Pan-STARRS1 (PS1) surveys were conducted with a 7 deg field-of-view camera mounted on a 1.8 m telescope near theHaleakala summit, Hawaii, in the five broadband grizy SDSS-like filters (Chambers et al. 2016). The camera has a pixel sizeof 0.258 (cid:48)(cid:48) pix − (Tonry & Onaka 2009). PS1 includes both the3 π and Medium Deep surveys. The former was completed in2014 and made publicly available in DR1 and DR2. It cov-ers 30 000 deg on the Northern sky down to –30 deg in thefive grizy filters with a depth of 21–23 mag. The median see-ing FWHM is 1.31 (cid:48)(cid:48) , 1.19 (cid:48)(cid:48) , and 1.11 (cid:48)(cid:48) in g , r , and i bands, re-spectively, but reaches > (cid:48)(cid:48) , > (cid:48)(cid:48) , and > (cid:48)(cid:48) over 20% ofthe footprint (Chambers et al. 2016). The Medium Deep surveyconsists of 10 fields covering a total of 70 deg , with multiplevisits in five filters optimized for transient detection. The few
2. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search
Fig. 1.
Left:
Examples of strong gravitational lens systems mocked up by painting COSMOS lensed sources on PS1 stack images, and used aspositive examples for training.
Right:
PS1 postage stamps of a subset of the 90 000 galaxies used as negative examples, including face-on spirals,massive LRGs, and field galaxies with similar gri colors as the mocks. All postage stamps are 20 (cid:48)(cid:48) × (cid:48)(cid:48) . hundred exposures will eventually provide deep stacks with 5 σ point source detection limits down to i ∼ . π survey that overlaps nicely with on-going optical time-domain surveys on the Northern hemisphere (e.g., ZTF, Masciet al. 2019) which provide a wealth of astronomical transients,including strongly-lensed SNe. PS1 extends the SDSS to lowerdeclinations and achieves higher depth. In particular, we applyour pipeline to gri stack images from DR1 which provide theoptimal coaddition of individual exposures and have higher 5 σ point-source sensitivities than z and y , probing down to 23.3,23.2, and 23.1 mag in g , r , and i , respectively (Chambers et al.2016). These three bands conveniently span a wavelength rangesensitive to young stellar populations in blue star-forming galax-ies at z (cid:38) π im-ages have non uniform coverage on small scales due to the stack-ing process. We found that limiting magnitudes vary by up to ∼ ff ect in our analysis.
3. Overview of the lens-search method
In this paper, we aim at identifying galaxy-scale strong lensingsystems on the extragalactic sky covered by PS1. We focus oursearch on typical high-redshift galaxies strongly lensed by mas-sive LRGs, which have a higher lensing cross-section (Turneret al. 1984) and smooth light profiles that help separate theforeground and background emissions. In particular, given thelong standing di ffi culty in distinguishing strong lensing featuresfrom arms of low redshift spirals, lenticular galaxies, tidal tailsand other contaminants with arc-like features (e.g., Huang et al.2019; Jacobs et al. 2019a), restricting to lens LRGs increases our chance of robustly identifying multiple lensed images with the > (cid:48)(cid:48) average PSF FWHM of PS1.Selecting these rare systems on the entire Northern sky re-quires an e ffi cient analysis of the properties of the 3 billionsources detected in the PS1 3 π survey image stacks. To cir-cumvent memory limitations, a number of CNN searches inthe literature have focused on subsets of galaxies with LRG-like photometry using, for instance, the Baryon OscillationSpectroscopic Survey (BOSS) sample (Schlegel et al. 2009;Dawson et al. 2013) or dedicated color / magnitude gri cutsadapted from Eisenstein et al. (2001). However, this approachrequires low contamination from lensed images to the lens multi-band photometry and was essentially applied to deeper sur-veys with better image quality (sub-arsec PSF FWHM in op-tical bands) than Pan-STARRS, such as the Hyper Suprime-Cam Subaru Strategic Program (HSC-SSP, Aihara et al. 2018;Sonnenfeld et al. 2018) or the Kilo-Degree Survey withOmegaCAM on the VLT Survey Telescope (KiDS, de Jong et al.2013; Petrillo et al. 2019). Due to the lower image quality, ap-plying such simple cuts on the photometry tabulated in Pan-STARRS DR2 catalogs would exclude significant fractions ofinteresting systems with strongly lensed arcs blended with thelens and altering its photometry. We therefore adopt a two-stepapproach: (1) a catalog-based neural network classification ofsource photometry, (2) a CNN trained on gri image cutouts.In addition, we aim at finding wide-separation lens systemsbecause these configurations provide longer time delays of afew days to weeks between multiple images, which is crucialto measure accurate, microlensing-free time delays for cosmol-ogy (Huber et al. 2019). Recently, the extensive follow-up of thelensed Type Ia SN iPTF16geu at z = . θ E ∼ . (cid:48)(cid:48) (Goobaret al. 2017) illustrated the di ffi culty in reaching the time-delayprecision required for cosmography on small-separation systems( ∆ t < ff ective search
3. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search strategy for the Pan-STARRS survey with limited angular reso-lution, and will help triggering timely imaging and spectroscopicfollow-up.As further described in Sect. 5, CNNs capture image char-acteristics by learning the coe ffi cients of convolutional filters(kernels) of given width and height and creating a range of fea-ture maps. They are invariant to translation and rotation. Duringthe learning phase, the CNNs rely on training sets with repre-sentative labelled images to minimize the di ff erence betweenpredictions and ground truth. Classification algorithms requiretraining sets of a few 10 to a million of labelled images de-pending on the number of classes, image complexity and net-work depth. In contrast to the recent computer vision imagerecognition challenges using deep CNNs (e.g., Russakovskyet al. 2015), relatively modest training sets of few 10 exam-ples are su ffi cient for our two-class problem applied to small,galaxy-scale image cutouts (e.g., Jacobs et al. 2017, 2019a).Nonetheless, the small number and heterogeneous properties ofspectroscopically-confirmed strong lenses (see the MasterLensdatabase) make it necessary to use simulated systems.Generating realistic mocks that account for the complexityof PS1 stack images is a critical ingredient to reach optimalclassification performances (e.g., Lanusse et al. 2018). We con-struct our mocks by painting lensed arcs on PS1 gri images ofLRGs with known redshift and velocity dispersion from SDSSspectroscopy. This approach captures the 3 π survey properties,such as background artifacts, the presence of line-of-sight neigh-bouring galaxies, and local variations of seeing FWHM, ex-posure time and noise levels, while also accounting for varia-tions in individual bands. In contrast to fully-simulated images,using real cutouts also guarantees positive examples that bestmimic the small scale background properties inherited from thecomplex masking and stacking of individual PS1 exposures. Asbackground sources, we use representative high-redshift galax-ies from the COSMOS field. High S / N image cutouts are takenfrom HSC-SSP and Hubble Space Telescope (HST) to simulatelens distortions and magnifications, and we use gri bands (simi-lar filter set as PS1) to provide color information.In Sect. 4, we present the selection of lens and source galax-ies and the pipeline to produce a set of mocks. Our first catalog-level network, described in Sect. 5.2, is trained on the multibandphotometry of mocks and non-lens systems from the PS1 cata-log, and assigns an output score, p cat , ranging between 0 and 1. Amuch lower fraction of sources with p cat > . p CNN . We eventually examine visually all sources with highestscores to assign grades and collect a list of high-confidence can-didates for future validation.
4. Simulating galaxy-scale strong lenses
Realistic strong lensing simulations require knowledge on thelens mass distribution and redshift. We therefore draw our sam-ple of lens LRGs from the SDSS spectroscopic samples withreliable velocity dispersion measurements, to have a proxy ofthe lens total mass. We use the SDSS large scale structure cata-logs of galaxies and QSOs for cosmological studies, includingLOWZ and CMASS samples for BOSS (from SDSS DR12),and the higher-redshift LRG catalog for eBOSS (from SDSSDR14, Bautista et al. 2018). QSOs are excluded using SDSS class flag. This results in a broad sample of LRGs selectedfor their redder rest-frame colors using gri color / magnitude cuts (Eisenstein et al. 2001). The sample is volume limited up to z ∼ . . < z < . z med ∼ .
7, Prakash et al. 2016) due to a combination ofoptical / mid-infrared cuts in SDSS and WISE bands.We clean this spectroscopic catalog to keep LRGs with reli-able velocity dispersions, using v disp ≤
500 km s − and v disp , err ≤
100 km s − , and obtain 1 192 472 LRGs to build the mocks. Wethen cross-match with the PS1 catalog to obtain their photome-try, image depth and seeing FWHM in PS1. The sample of galaxies used to mock up high redshift lensedsources is drawn from the COSMOS field to take advantage ofthe wealth of existing data including ultra-deep optical imag-ing, multiband photometry, spectroscopic follow-up, and mor-phological classification. We select galaxies with morpholog-ical information from G alaxy Z oo : HST (Willett et al. 2017)and within the COSMOS2015 photometric catalog (Laigle et al.2016) that also lists physical parameters from SED fitting. Theformer is a citizen science project that extends the originalGalaxy Zoo (Lintott et al. 2008, 2011; Willett et al. 2013) witha thorough visual classification of galaxies with ACS imag-ing from the Hubble legacy surveys (see Scoville et al. 2007;Koekemoer et al. 2007, for the COSMOS field). In particular forCOSMOS, G alaxy Z oo : HST relies on 3-color images obtainedby combining the HST F814W mosaic with color gradients fromground-based imaging .We clean the resulting catalog from sources identified asstars or artifacts (COSMOS2015 flag or visual identification),and remove very extended galaxies with R e ff > . (cid:48)(cid:48) , as well asgalaxies contaminated by emission from companions within 5 (cid:48)(cid:48) ,and brighter by 1 mag in r band (Laigle et al. 2016). The out-put sample includes 52 696 galaxies for the strong lensing simu-lations. Redshifts are taken from public spectroscopic redshiftcatalogs drawn from surveys with VLT / VIMOS (zCOSMOS-bright, Lilly et al. 2007), VLT / FORS2 (Comparat et al. 2015),Subaru / FMOS (Silverman et al. 2015), VLT / VIMOS (VUDS, LeF`evre et al. 2015; Tasca et al. 2017), Keck / DEIMOS (Hasingeret al. 2018), or the best photometric redshift estimate from Laigleet al. (2016) for galaxies without z spec available.For the purpose of using this pipeline in future lensed SNsearches, the properties of COSMOS sources are compared withexpectations for high-redshift SN hosts. Firstly, the cleaned cat-alog has a redshift distribution peaking at z ∼ . z (cid:38) .
5, akin to the mock lensed SN catalog of Oguri& Marshall (2010), both for LSST-like imaging or for current,shallower surveys probing down to R ∼ > ∼
13% of SNe of all types at z (cid:46) . − r / r − J cuts, with Galaxy Zoo morphologies to conclude that our sam-ple is strongly dominated by star-forming galaxies, with ∼ Galaxy Zoo: CANDELS (Simmons et al. 2017) performs similarclassifications using deeper, multiband HST images but only over a sub-set of the COSMOS field (Grogin et al. 2011; Koekemoer et al. 2011)and strongly restricts the sample size.4. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search classified as quiescent. This shows that our sample of sourcesbroadly matches the expected properties of SN hosts.
For the lenses, PS1 gri image cutouts of 20 (cid:48)(cid:48) × (cid:48)(cid:48) are down-loaded from the PS1 cutout service . We characterized the imagedepth in individual cutouts with SExtractor (Bertin & Arnouts1996) and verified that the observing strategy leads to nearlyuniform depth. Although it depends on several observing fac-tors, the depth is weakly correlated with the number of indi-vidual warp exposures used in a given stack. In particular, itrapidly drops by 0.2–0.3 mag for the 10% of cutouts obtainedby coadding less than 8, 10, and 12 frames in g , r , and i bands,respectively. A small fraction of <
5% of these PS1 images werediscarded from the analysis.Multiband images of COSMOS galaxies used as lensedsources are taken from the first data release of the HSC SSP(Aihara et al. 2018). The HSC ultra-deep stacks are providing thedeepest optical exposures with best image quality over the 2 deg of COSMOS and are well suited for the simulation pipeline. The5 σ point-source sensitivities are 27.8, 27.7, 27.6 mag, in g , r , i , respectively, and seeing conditions are excellent, with medianvalues of 0.92 (cid:48)(cid:48) , 0.57 (cid:48)(cid:48) , and 0.63 (cid:48)(cid:48) in g , r , and i bands, and negli-gible variations over the COSMOS field (Tanaka et al. 2017).We download gri cutouts of 10 (cid:48)(cid:48) × (cid:48)(cid:48) , su ffi cient to encloseall emission from galaxies with R e ff < . (cid:48)(cid:48) . Fainter companionswithin a few arcsec are masked using segmentation maps createdin r band with SExtractor (using relatively few deblending sub-thresholds, Bertin & Arnouts 1996) to isolate the central galaxyof interest. To overcome the limited spatial resolution of ground-based images, we combine these frames with the HST F814Whigh-resolution images over the COSMOS field (see Leauthaudet al. 2007; Scoville et al. 2007; Koekemoer et al. 2007) toproduce pseudo color images following the steps described inGri ffi th et al. (2012). First of all, F814W images are aligned andrescaled as if observed in HSC i band, and masked HSC framesare resampled with SWarp (Bertin et al. 2002) to the HST scalingof 0.03 (cid:48)(cid:48) pix − using nearest-neighbour interpolation. Secondly,we multiply each resampled frame by an illumination map, de-fined as F814W divided by HSC i band. This process preservesHSC source photometry, and results in gri images with high-resolution light profiles and colour gradients with seeing-limitedresolution. Due to the limited angular resolution of Pan-STARRS, we fo-cus on the search for wide-separation lens systems with brightarcs that can be easily recognized by eye. We impose a lowerlimit on the Einstein radius θ E of mocks of 1.5 (cid:48)(cid:48) , larger than themedian FWHM of PS1 seeing in gri bands. This ensures that in-dividual counter-images are well deblended from each other inthe mocks (albeit often blended with the lens). Each lens deflec-tor drawn from the LRG catalog is cross-matched with a ran-dom COSMOS source at z source > z deflector , rejecting pairs with θ E < . (cid:48)(cid:48) , and repeating the process iteratively to obtain 90 000lens + source pairs. Focusing on larger Einstein radii amountsto selecting LRGs in the high-mass range, with v disp ∼ − , and redshifts z d ∼ . http: // hla.stsci.edu / fitscutcgi interface.html Calculated given the known lens redshift and velocity dispersionand the known source redshift. put BOSS sample, and sources in the redshift range z s ∼ . θ E range of 1.5–3.0 (cid:48)(cid:48) , which isdominated by galaxy-scale dark-matter lens halos on the high-end of the mass distribution (Oguri 2006), while group-scalelenses contribute predominantly for image separations above 3 (cid:48)(cid:48) . θ E values are not uniformly distributed but drop by a factor 100from 1.5 (cid:48)(cid:48) to 3.0 (cid:48)(cid:48) , akin to real galaxy-scale lenses, implying thatour CNN will be predominantly exposed to mock systems with θ E ∼ . (cid:48)(cid:48) .For each pair, mock images are created with the simulationpipeline described in Schuldt et al. (in prep.). In short, the lenspotential is modeled with a Singular Isothermal Ellipsoid (SIE)profile, based on the known v disp and z d , and using the cen-troid, axis ratio, and position angle from the i -band light distribu-tion, with random perturbations typical of SLACS lenses (Boltonet al. 2008). The combined HSC + F814W cutouts of COSMOSsources are randomly positioned in the source plane, over re-gions next to the caustics corresponding to magnifications µ ≥ GLEE software (Suyu & Halkola 2010; Suyu et al. 2012). The result-ing frames are convolved with the Pan-STARRS PSF model de-scribed below, resampled and rescaled using Pan-STARRS zero-points, and eventually coadded with the lens LRG cutouts to ob-tain the final mock image. The process is repeated for gri bands.In order to produce a set of mocks with systematically brightlensing features, we artificially boost the lensed source bright-ness by one magnitude in all bands. Systems with lensed sourceshaving S / N <
10 in i band are placed iteratively closer to thecaustics to meet this S / N threshold, or completely discarded.The Pan-STARRS analysis pipeline computes PSF modelsat individual positions of stack images over a grid of about 8 (cid:48) steps, and interpolates these models to predict the PSF FWHMacross the sky (Magnier et al. 2016). This introduces devia-tions between the modeled and true PSF, since the later varieson very small scales due to stacking of individual exposureswith variable FWHM. However, we found that these deviationsare usually within 10% when comparing the tabulated FWHMswith those measured on isolated, unsaturated stars from GSC-DR2 (Lasker et al. 2008). We therefore create a library of gri
PSF models in steps of 0.05 (cid:48)(cid:48)
FWHM, by stacking PS1 postagestamps of 9 to 11 stars with adequate PSF FWHM. For eachmock, lensed arcs are convolved with the PSF model corre-sponding to the PSF FWHM listed on PS1 tables at the positionof the lens LRG.We generate a total of 90 000 mock lens systems (see Fig. 1).Lens LRGs selected multiple times are rotated by k π/ ff erent lensed arc con-figurations, so the networks never get the exact same image sev-eral times as input. The mocks cover realistic source colors andlensing configurations, including quads, near-complete Einsteinrings, fold and cusp arcs, and doubles (see Schuldt et al., inprep.). Constraining the source plane positions to large magnifi-cations likely biases our set to lower fractions of doubles than inblind samples of real lenses. Focusing the CNN lens search on a subset of the 3 billion sourcesdetected in the PS1 3 π survey requires a pre-selection of sourcesbased on their properties released in public catalogs. For thispurpose, we compute the photometry of mock lensed systemsin the same way as the PS1 image processing pipeline (Magnier Sources fainter than 10 times the local background rms level. 5. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search
Fig. 2.
Aperture magnitudes and colors of galaxies in the lens LRG catalog (red regions) and in the set of 90 000 mock lens systems (blue regions).The red and blue dots show the median of distributions.
Left: ( g − i ) vs. i diagram for the R3 aperture, a circular aperture of 1.04 (cid:48)(cid:48) radius. Right: Di ff erence in ( g − i ) color between the inner R3 aperture and concentric annuli between R3 and R4 (1.76 (cid:48)(cid:48) outer radius), and between R4 andR5 (3.00 (cid:48)(cid:48) outer radius). Grey regions mark the position of 100 000 random sources with reliable gri aperture photometry selected from the PS1DR2 catalogs. Orange regions show galaxies with p cat > . σ contours. et al. 2016). Fixed aperture photometry is particularly importantto measure reliable colors. We derive the integrated magnitudesof our mocks within the four smaller PS1 circular apertures of1.04 (cid:48)(cid:48) (R3), 1.76 (cid:48)(cid:48) (R4), 3.00 (cid:48)(cid:48) (R5), and 4.64 (cid:48)(cid:48) (R6) radii, whichare best suited to capture color gradients due to the presenceof lensed arcs at angular separations of 1.5–3.0 (cid:48)(cid:48) with respectto the lens center. The two largest of these apertures are alsoexpected to be relatively good proxies of the integrated magni-tudes of mocks. We use SExtractor in dual-image mode, with a3 σ detection threshold in r-band, and assuming an ideal sky sub-traction in PS1 deep stacks (see details in Waters et al. 2016). Tocompute the aperture magnitudes of a given mock, the imagezero-points are taken from the PS1 catalog of stack detections atthe position of the LRG used to produce this mock.The method is tested on LRG-only images by comparingSExtractor estimates with those from the PS1 catalog, for stan-dard stacks. For the four apertures, fitting the distributions ofmagnitude o ff sets between these two estimates with Gaussianfunctions leads to µ = . σ = . g band, µ = . σ = . r band, µ = .
00 and σ = . i band, which proves the overall robustness of ourphotometry. Unsurprisingly, the scatter only rises above theseaverage values for the large aperture magnitudes of fainter ob-jects, which mostly enclose background noise. These residualbiases up to 0.1–0.2 mag for (cid:38) ff erences in the local background subtraction between bothmethods, or the contamination from neighbours which are sub-tracted by the PS1 pipeline (Magnier et al. 2016) but not withSExtractor. Nonetheless, these o ff sets remain minor comparedto other uncertainties in the analysis. On the contrary, Kron,Petrosian and Sersic photometry are discarded due to systematicbiases with respect to values tabulated in PS1 catalogs.
5. Systematic search of strong lenses
The next sections describe the steps followed for the generic lenssearch on the full Pan-STARRS 3 π survey. As shown in Fig. 2, the 90 000 mock lens systems have glob-ally bluer colors than the LRG sample due to the relative colorof lensed arcs, and they are brighter than ∼
70% of sources de-tected in the PS1 stack images. This population of fainter andbluer galaxies can be excluded from the analysis. We use simplecolor-magnitude cuts in the ( g − i ) vs. i , ( g − r ) vs. r , and ( r − i )vs. i diagrams for the R3, R4, R5, R6 circular apertures to ruleout regions in these diagrams that are not representative of themocks (i.e. not colored in blue in Fig. 2). These cuts are conser-vative and include 96% of the mocks, according to the aperturemagnitudes obtained in Sect. 4.5, while excluding ∼
84% of PS1sources. They are applied to the complete catalog of stack detec-tions from PS1 DR2 using the PanSTARRS1 Catalog ArchiveServer Jobs System , with detectionFlags3 > r -band cuts from Farrow et al. (2014): r Kron − r PSF < − . + . × ( r Kron − + . × ( r Kron − .This selection conservatively includes 98% galaxies, thosediscarded being mainly at the faint end ( r (cid:38)
21) and with highermagnitudes than our mocks. While these cuts misidentify sat-urated stars with r (cid:46)
14 as galaxies (Farrow et al. 2014), suchbright sources have already been excluded from the analysis.Regions with elevated Galactic dust extinction are removedas strong reddening could alter our selection on the catalog level.The interstellar dust reddening 2D map of Schlegel et al. (1998)are loaded with the dustmaps python interface from M. Green(2018). After converting to PS1 bandpasses using coe ffi cients inTable 6 of Schlafly & Finkbeiner (2011), we apply a reddeningthreshold of E( g − i ) < .
3. These steps result in a catalog of 23.1million galaxies for classification. http: // casjobs.sdss.org / CasJobs6. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search
Limitations due to download speed of PS1 cutouts fromthe archive can be overcome by further reducing the sizeof this catalog with additional selection criteria. The simplecolor / magnitude cuts applied to the complete PS1 catalog donot capture all photometric properties of mock lens systems,such as their precise locus on two-dimensional color / color andcolor / magnitude diagrams or their radial color gradients. For in-stance, Figure 2 indicates that mocks are generally redder withinthe smaller R3 aperture of 1.04 (cid:48)(cid:48) radius than within external,concentric annuli between R3 and R4, and between R4 and R5.These gradients are caused by the presence of bluer, lensed arcsat > (cid:48)(cid:48) from the lens center and disappear on the LRG-onlysample. To exploit this information, we train a simple fully-connected neural network on the photometry of mocks and ran-dom PS1 sources using aperture fluxes that ensure robust colors.The data set contains gri fluxes in the four apertures for90 000 lens and 90 000 non-lens examples as inputs, and theground truth labels of 1.0 and 0.0, respectively, as outputs forbinary classification. The negative examples are fluxes of ran-dom sources that match our loose color / magnitude gri cuts inSect. 5.1. The data set is split into training, validation, and testsets with respective fractions of 56%, 14%, and 30%. All fluxesare normalized to the average over the entire data set in orderto speed up the learning process. The network architecture con-sists of 12 dimensional input data, three fully connected hid-den layers of 50, 30, and 5 neurons each, with Rectified LinearUnit (ReLU, Nair & Hinton 2010) non-linear activations , and asingle-neuron output layer with sigmoid activation .During the training phase, the network derives a model forclassifying galaxies in the training set as lenses or non-lensesaccording to their input photometry, as briefly summarized withthe following stages. After the weight parameters and bias ineach neuron are initialized, a subset of the training data is passedthrough the entire network to calculate predicted labels (forwardpropagation), and the di ff erence between predictions and groundtruth labels is quantified with a loss function L . This informationis propagated to the network weights and biases (back propa-gation, Rumelhart et al. 1986) which are then modified usinga gradient descent algorithm to minimize the total loss and im-prove the model. These stages are repeated iteratively to performa complete pass through the entire training set, corresponding toone epoch, and then over multiple epochs until the model reachesoptimal accuracy. After each epoch, the validation loss is eval-uated by classifying inputs from the validation set, in order todetermine whether the decrease in training loss reveals betterperformance or an overfitting to the training set. After training,the network performance is finally quantified using independentdata from the test set. Further details can be found in the reviewof LeCun et al. (2015).Network parameter optimization is performed via mini-batchstochastic gradient descent, a common variant that consists ofsplitting the training set into small batches and adjusting theweights according to the average corrections over each batch.Our network minimizes the cross-entropy loss function which ReLU(x) = max(x , sigmoid( x ) = + e − x ) Fig. 3.
Architecture of the convolutional neural network, inspired fromLeNet (LeCun et al. 1998), and comprised of three convolutional layerswith 11 ×
11, 7 ×
7, and 3 × × = penalizes robust and incorrect predictions, and is expressed asfollows for a binary classification problem L ( y , p ) = − N N (cid:88) i = y i log( p i ) + (1 − y i ) log(1 − p i ) (1)where y i are the ground truth labels, and p i the network predic-tions, namely scores in the range [0 ,
1] resulting from the sig-moid activation on the output layer. The loss is computed overeach batch of size N . To avoid unbalanced splits of the data set,we use 5-fold cross-validation that consists of reshu ffl ing thetraining and validation sets and building the performance met-rics. Cross-validation runs trained over 500 epochs are used tooptimize the neural network hyperparameters with a grid search,varying the learning rate over the range [0 . , .
1] and theweight decay over [0 . , . p cat > . gri aperture pho-tometry is consistent with the mocks. Fig. 2 illustrates the net-work predictions by comparing the colors and magnitudes ofrandom galaxies, mocks, and galaxies having p cat > .
5. Thegood agreement between 2 σ contours of mocks and p cat > . / magnitude diagrams, as well as their color variationswithin di ff erent apertures. Moreover, our photometric selectionis successfully tested on the aperture fluxes of known stronglensing systems listed as grade A or B with θ E = . (cid:48)(cid:48) inthe MasterLens database , and with lensed arcs visible in thePS1 stack images. We therefore keep these 1 050 207 galaxiesfor CNN classification. Data sets for the CNNs include 180 000 images in g , r , and i bands with the same fraction of positive (lens) and negative (non-lens) examples. Positive examples are taken from the sample ofsimulated lens galaxies with θ E > . (cid:48)(cid:48) described in Sect. 4.4.The choice of negative examples is strongly influencing the net-work predictions and is modified iteratively to improve the net-work performances (see Sect. 5.5). In short, we boost the frac-tion of galaxies with specific morphological types using incor-rect identifications of strong lenses in previous networks. This http: // admin.masterlens.org 7. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search Fig. 4.
Training of the CNN with optimized hyperparameters and usingearly stopping. The training loss (red curve) and validation loss (bluecurve) are taken as the average of all cross-validation runs (light redand blue curves).
Fig. 5.
Distribution of network predictions compared with the ground-truth for lenses (green) and non-lenses (orange) in the test set. allows the network to learn how to discriminate strongly lensedarcs from the usual contaminants such as extended arms of lowredshift spirals, lenticular galaxies, and mergers, and to distin-guish isolated LRGs from LRGs with the relevant strong lensingfeatures depicted in the mocks. Our resulting set of 90 000 neg-ative examples includes: –
30% LRGs selected directly from the catalog of SDSS LRGson the high-end of the mass distribution used to create themocks, –
20% spirals classified as likely face-on galaxies in GalaxyZoo 2 (Willett et al. 2013) and with r -band Sersic radii < (cid:48)(cid:48) from PS1 (Flewelling et al. 2016), to restrict to blue spiralarms with similar extension as the lensing features present inthe mocks, –
10% smooth, isolated galaxies from Galaxy Zoo 2 withoutbright companion and bluer colors than LRGs, – <
1% galaxies with apparent dust lanes identified in GalaxyZoo 2 (Willett et al. 2013), –
32% randomly selected galaxies from the PS1 catalog, in-cluding diverse types, groups and mergers, and with negligi-ble contamination from the rare strong lenses, –
7% false positives from previous neural networks selected byvisually classifying candidates with scores > p cat > . gri bands. Some examplesare shown in Fig. 1. The data set is split into training, validation,and test sets with the same proportions as before (Sect. 5.2).We use data augmentation and apply random shifts between–5 and + ff sets. This results in input images of 70 × k π/ × × ×
11, 7 ×
7, and 3 × × = gri image cutouts,while reducing the network parameters. Dropout regularization(Srivastava et al. 2014) with a dropout rate of 0.5 is applied be-fore the fully connected layers. This is an e ffi cient regularizationmethod that consists of randomly ignoring neurons during train-ing in order to reduce overfitting on the training set and improvethe CNN generalization. The output layer consists of a singleneuron with sigmoid activation and results in a score, p CNN , inrange [0 ,
1] which corresponds to the network lens / non-lens pre-diction . Our CNN with moderate depth is well suited for binaryclassification of small PS1 cutouts.During the training process, the CNN learns the relevantpatterns in gri images by adjusting the convolutional kernelweights, through a minimization of the binary cross-entropy lossbetween ground truth and predicted labels. After the gradientcalculation and optimization, information learned by the net-work is stored in the two-dimensional filters. As for the catalog-level network, we use mini-batch gradient descent with a batchsize of 128 and performed five cross-validation runs. We find anoptimal learning rate and weight decay of 0.0006 and 0.001, re-spectively, using a grid search with momentum fixed to 0.9. Thenumber of training epochs is then chosen from the minimum av-erage validation loss over the cross-validation runs, which cor-responds to optimal network performance without overfitting.The evolution of the training and validation loss for the net-work with optimized hyperparameters is shown in Fig. 4, untilepoch 47 which corresponds to the lowest validation loss. Thegap between both curves (generalization gap) is small, show-ing that the model predictions do not deteriorate much on newdata with similar properties as the training set. The final networkperformances are characterized with the test set which was not For uncalibrated networks, this score di ff ers from the likelihood ofcorrect classification.8. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search Fig. 6.
Receiver Operating Characteristic curve for the trained CNNshowing the true position rate (TPR) as function of the false positiverate (FPR) for di ff erent lens identification thresholds. The correspond-ing area under curve (AUC) is 0.985. seen during training and validation and contains about 54 000entries. In Fig. 5, we show the model probability predictions forall lens and non-lens examples in the test set. Lenses dominatethe distribution for p CNN > .
6. The model reaches 94.2% ac-curacy, 93.1% purity and 95.5% completeness on this set sug-gesting good pattern recognition abilities on new images . Inaddition, the Receiver Operating Characteristic (ROC) curve inFig. 6 illustrates the relation between the true positive rate (TPR,the number of lenses correctly identified over the total numberof lenses) and the false positive rate (FPR, the number of non-lenses identified as lenses over the total number of non-lenses) inour trained model. The curve is obtained by varying the networkprobability threshold for lens identification in the range [0 , = . p CNN > .
5, or 77.0% of lenses in the test set with a FPR of 0.8%using p CNN > . p cat > . The performance of the network is further evaluated on galaxieswith various properties on the test set. Fig. 7 depicts the nor-malized distributions of scores for positive examples in di ff erentbins of θ E , θ E / R e ff , and r Kron , where R e ff is the e ff ective radiusfrom the r -band Sersic fit of the LRG-only image (Flewellinget al. 2016). Somewhat counterintuitively, scores are closer to1.0 for lenses with θ E < (cid:48)(cid:48) . The low fraction of θ E > (cid:48)(cid:48) in thetraining set might explain the slightly lower CNN performanceon these systems, which generally do not have the most chal-lenging morphologies. Histograms for θ E / R e ff demonstrate theability of the network to assign p CNN closer to 1.0 for mocks Accuracy is defined as the sum of true positives and true negativesover the total number of systems (lenses + non-lenses), and purity isdefined as the number of lenses correctly identified with p CNN > . p CNN > . with Einstein radius larger than the e ff ective radius of the lenslight distribution, where lensed arcs are in principle better de-blended from the lens. We also find a higher fraction of scores p CNN > . r Kron (i.e. fainter LRGs),perhaps because the brightest lenses outshine the lensed sourceemission. Nonetheless, these variations remain generally minor.Finding acceptable network performances on the test setmight be misleading as it relies on our choices in simulatingthe strong lenses and assembling a set of negative examples.A valuable independent test consists of applying the CNN tostrong lenses from the literature. For that purpose, we collectall grade A and B galaxy-scale lenses in the MasterLens cat-alog, restricting to Einstein radii ∼ (cid:48)(cid:48) similar to the rangeprobed by our 90 000 mocks. These systems were discoveredfrom various techniques including the identification of emissionlines from star-forming galaxies behind LRGs using spatially-integrated spectra (SLACS, Bolton et al. 2008), and the analy-sis of high-quality imaging from HST (e.g., in COSMOS, Faureet al. 2008) or deep multiband surveys (e.g., CFHTLS SL2S,Cabanac et al. 2007; Gavazzi et al. 2014; More et al. 2016). Mostof these lenses are not detectable in gri Pan-STARRS stacks andneed to be excluded from our test set. We thoroughly scannedall PS1 gri single-band and 3-color images of this sample to findthose with detected lensed arcs, and assembled a test set of 16systems. While these published lenses have colors, z d , z s , andconfigurations similar to our mocks, some of their multiple im-ages are strongly blended with the lenses and di ffi cult to identifywith PS1 data.Pan-STARRS cutouts of these 16 lenses are scored withour trained neural network and the results are presented inFig. 8. A total of 14 /
16 lenses are correctly identified as p CNN > . /
16 and 7 /
16 have higher scores p CNN > . p CNN > .
9, respectively. SL2SJ0217 − + − z d = . + p CNN = . − / N, blue counter image on the other sideof the lens galaxy. SDSSJ1112 + + + (cid:48)(cid:48) cuto ff applied to our simulations. The CNN per-formance in this regime remains acceptable, as illustrated bythe scores of 0.786, 0.612, and 0.971 assigned respectively toSDSSJ1134 + + − θ E ∼ . (cid:48)(cid:48) . In contrast, with θ E = . (cid:48)(cid:48) due to its group-scale environment (Auger et al. 2013), CSWA21 falls on the up-per range of the Einstein radius distribution that is underrepre-sented with only a few hundred examples in our training set. Thissystem is nonetheless given a score of 0.873 and confirms thatthe CNN can identify these simple configurations. Interestingly,the four systems with ≥ p CNN > .
9. The small number of testlenses however prevents robust estimates of the method purityand completeness.
9. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search
Fig. 7.
Normalized distributions of CNN scores for mocks lenses included as positive examples in the test set, for di ff erent ranges of Einstein radii(left), θ E / R e ff (middle), and r Kron (right).
Fig. 8.
Three-color images of the 16 confirmed lens systems in theMasterLens database that have clear strong lensing signatures in Pan-STARRS images, and Einstein radii of 1–3 (cid:48)(cid:48) similar to our mocks.The CNN scores are displayed at the top of each panel. Images are15 (cid:48)(cid:48) × (cid:48)(cid:48) . The final version of the CNN from Sect. 5.3 is selected from arange of networks with di ff erent architectures, after testing theimpact of the training set content. To identify the optimal net-work, we compare scores assigned to the 16 known test lensesand require low false positive rates by examining gri cutoutsof the few hundred galaxies with highest scores p CNN . Overall,di ff erent choices of positive and negative examples have muchstronger impact than changes in the CNN architecture.The sets of negative examples tested include: (1) random PS1sources drawn from the pre-selection in Sect. 5.1; (2) typicalLRGs selected as in Eisenstein et al. (2001), mostly less massivethan LRGs in mocks; (3) high-mass LRGs similar to those usedin Sect. 4; (4) a combination of LRGs, face-on spirals, and ran-dom sources (varying the fractions of LRGs and contaminants).Scores on the MasterLens systems from CNNs using sets (1)and (4) are comparable to those in Fig. 8, but introduce an over-whelming number of (cid:38)
400 000 galaxies with p CNN > . (cid:38)
250 000 with p CNN > . θ E lower limit, and (2) suppressing the artifi-cial boost in arc brightness to get more realistic lens / source fluxratios. Both CNNs trained on fainter arcs, more strongly blendedwith the lens significantly reduce the fraction of genuine lens ex-amples with scores p CNN ∼ ff e & Szegedy 2015). Eachof these changes degrade the network performance as measuredfrom the loss and ROC curves on the test set, and from the 16MasterLens systems. Using dropout normalization before fullyconnected layers (as in Fig. 3) turns out to be the most e ffi cientsolution to reduce overfitting. Galaxies with high CNN scores, p CNN , are visually inspected bydi ff erent authors to assign a final grade. We started classifyinggalaxies with p CNN close to 1.0, and progressively lowered thethreshold to introduce additional galaxies until the fraction ofreliable candidates from visual inspection became too low.We grade the single-band and 3-color PS1 cutouts in gri bands, zoomed to 12 (cid:48)(cid:48) × (cid:48)(cid:48) , and optimally displayed with lin-ear and arcsinh scaling using dedicated scripts to emphasizefaint, sometimes blended strong lensing features . To aid the vi-sual classification, we plot grz single-band and 3-color postagestamps from the DESI Legacy Imaging Surveys that signifi-cantly overlaps the PS1 3 π survey footprint on the extragalacticNorthern sky, and that provides slightly deeper, higher qualityimages (Dey et al. 2019), and we plot residual frames from sub-traction of the best-fit light profile. On smaller regions of the sky,we also include gri images from HSC DR2 wide-field surveys(Aihara et al. 2019). The set of CNN candidates is divided intofour equal parts, each inspected either by R. C., S. S., S. H. S., orS. T., in order to assign one of the following grades : : non-lens, : maybe a lens, : probable lens, : definite lens, similarly toSonnenfeld et al. (2018) and Jacobs et al. (2019a). All candi-dates with grades ≥ ≥ code adapted from https: // github.com / esavary / Visualisation-tool10. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search
Fig. 9.
Top:
Pan-STARRS 3-color gri images of a subset of candidates with grades ≥ p CNN > . p CNN , left) and average visual grades(right). Candidates with PS1 names marked in orange have been previously published in the literature (see Table 1 and Sect. 6.1) Those marked inblue show unambiguous spectral signatures of high-redshift background sources in our inspection of SDSS BOSS DR16 data.
Bottom:
Examplesof random false positives with p CNN =
1. All cutouts are 15 (cid:48)(cid:48) × (cid:48)(cid:48) . Non-lenses are galaxies clearly identified as nearby spirals,ring galaxies, groups, or other contaminants from their morphol-ogy, or cutouts with artifacts. Candidates listed as grade 1 havefaint companions and / or weakly distorted features suggestingpossible strong lensing signatures, but may also correspond togalaxy satellites or spiral arms. Probable lenses show multipleelongated sources with similar colors, and orientation and angu-lar separation expected for counter-images, while the available3-color images cannot firmly rule out contaminants. Those as-signed grades of 3 have similar, although brighter, non-blendedand unambiguous signatures of galaxy-scale strong lenses.
6. Results and discussion
Out of the 1.1 million galaxies the CNN scores 598 130 with p CNN =
0, and 105 760, 12 382, and 1714 as candidate lenseswith p CNN > . p CNN > .
9, and p CNN =
1, respectively. Scores p CNN > . p CNN > . p CNN threshold which impacts the purity and complete-ness in a way that is di ffi cult to quantify. Predictions for mocklenses on the test set (see Fig. 5) and for known lenses (seeFig. 8) suggest that the majority of good candidates have scores
11. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens searchName RA Dec p CNN
Grade g Kron r Kron i Kron g aper r aper i aper Redshift NotesPS1J2226 + + ( ∗ ) (a), (b), (c)PS1J1821 + + + + + + ( ∗ ) PS1J1508 − − − − + + ( ∗ ) PS1J1322 − − + + ( ∗ ) PS1J0919 + + ( ∗ ) (c), (l)PS1J0353 − − − − − − − − + + ( ∗ ) PS1J2336 − − ( ∗ ) PS1J2248 − − ( ∗ ) (e)PS1J2247 + + ( ∗ ) PS1J2233 + + ( ∗ ) PS1J2202 + + − − − − + + + + + + + + ( ∗ ) (c)PS1J1553 − − + + ( ∗ ) PS1J1439 + + ( ∗ ) PS1J1422 + + ( ∗ ) (e)PS1J1411 + + ( ∗ ) PS1J1349 + + ( ∗ ) PS1J1241 + + + + + + ( ∗ ) PS1J0921 + + ( ∗ ) (f), (l)PS1J0907 + + ( ∗ ) PS1J0737 + + ( ∗ ) PS1J0717 + + − − Table 1.
Final list of galaxy-scale strong lens candidates with lens LRGs from our systematic search in Pan-STARRS (the complete table isavailable in Appendix A). These systems are selected as high confidence candidates with CNN scores > .
9, and average grades ≥ . g -, r − , and i -band Kron magnitudes of the lens and source blends from the PS1 catalog; g -, r − , and i -band aperture magnitudes of 1.04 (cid:48)(cid:48) radii coveringthe lens central regions; SDSS photometric redshifts or spectroscopic redshifts marked as ( ∗ ) where available; previously published confirmed orcandidate systems (grades A and B or equivalent). References are the following: (a) Jacobs et al. (2019a), (b) Diehl et al. (2017), (c) Sonnenfeldet al. (2018), (d) Huang et al. (2019), (e) Wong et al. (2018), (f) Petrillo et al. (2019), (g) Stark et al. (2013), (h) Auger et al. (2009), (i) Jacobset al. (2019b), (j) Lemon et al. (2019), (k) Wang et al. (2017), (l) Jaelani et al. (2020), and (m) Schirmer et al. (2010). p CNN > . . < p CNN < . p CNN > . ≥ (cid:46)
1% when extending to 0 . < p CNN < .
9, we restricted our final classification to p CNN > . ≥ . < p CNN ≤ .
00, 0 . < p CNN ≤ .
99 and 0 . < p CNN ≤ .
95, which demon-strates that the CNN learns meaningful information and assignshigh scores for most of the probable or definite lenses. The rate of grades ≥ p CNN > . π sur-vey. Finally, 37 additional candidates from previous CNNs with p CNN = . ≥
12. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search
Fig. 10.
Pan-STARRS lens candidate with prominent Balmer absorption features from a background galaxy at z s = .
185 blended with the BOSSspectrum of the lens LRG at z d = . (cid:48)(cid:48) diameter and the 1 σ noise level,respectively, and the blue line corresponds to the best-fit LRG template from the automated SDSS pipeline at λ < z d = . z s = .
185 (orange curve) correctly fits the spectrum over the entire range (see details in text). ing the total number of resulting lens candidates to 358 from ourtwo-step search.Examples of good candidates and false positives (grades ≤ / near-infrared imaging fromDES (Diehl et al. 2017; Jacobs et al. 2019b,a), CFHTLS (Moreet al. 2016; Jacobs et al. 2017), KiDS (Petrillo et al. 2017,2019; Li et al. 2020), DESI (Huang et al. 2019), and HSC(Wong et al. 2018; Sonnenfeld et al. 2018; Jaelani et al. 2020;Sonnenfeld et al. 2020). Since our network may also be sensitiveto lensed quasars with colors and configurations similar to ourmock lenses, we also cross-matched with the all-sky databaseof ∼
220 confirmed lensed quasars in the literature (Lemon et al.2019, and references therein) , and with previously identifiedlensed quasars from HSC (Chan et al. 2019) and from PS1 (Rusuet al. 2019). For the candidates not included in those tables, wesearched in the SIMBAD Astronomical Database .To our knowledge, besides the test lenses from theMasterLens database, 23 of our 358 CNN candidates are al-ready listed in the literature and corresponding referencesare listed in Table 1. The vast majority of them are alsogalaxy-scale strong lens candidates from ground-based multi- https: // / ioa / research / lensedquasars / http: // simbad.u-strasbg.fr / simbad / sim-fcoo band imaging searches and lack both spectroscopic and high-resolution imaging follow-up. The only galaxy-galaxy lens sys-tems confirmed with spectroscopy and / or space-based imag-ing are PS1J0143 + − − − + z d = . z s = . θ E = . (cid:48)(cid:48) ) are very similar to our mock lenses,and the system was assigned p CNN = . − z d = . z s = . θ E = . (cid:48)(cid:48) , akin to our mocks. It has p CNN = . − z d = .
181 and z s = .
463 which was discovered and mod-eled by Auger et al. (2009) using SDSS spectroscopy and HSTimaging. It has an Einstein radius of 1.5 (cid:48)(cid:48) , higher than the major-ity of lenses in the SLACS sample, which explains its elevatedscores from the CNN ( p CNN = . − p CNN = .
000 and grade = z = .
26 only 8 (cid:48) from the MS0451 − z = .
54. The HST F814W frame clearly resolves thebackground source into an extended arc and a compact counterimage, corresponding to θ E = . (cid:48)(cid:48) . Most other galaxy-galaxylenses in the literature were missed by our selection due to lim-ited PS1 depth or lens configurations not represented in ourmocks (e.g., higher Einstein radii). Finally, two of these 23 pub-lished systems are confirmed lensed quasars. PS1J2350 + p CNN = . + p CNN = . p CNN and visual inspection grades. Given that some visual iden-tifications rely on Legacy imaging, which is sometimes deeperthan PS1, we also show Legacy 3-color images of our candidateslocated in the Legacy footprint in Appendix A.
13. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search
Fig. 11.
Example of a lens candidate with multiple emission lines from a high-redshift background galaxy overlaid on the SDSS BOSS spectrumof the foreground LRG. The black, grey and blue lines show the observed spectrum, the 1 σ noise level, and the best-fit SDSS template for theLRG, respectively. The spectrum is zoomed on the spectral features associated with the background line-emitter at z = . z = . We inspect SDSS BOSS spectra from the 16 th data release avail-able for 104 out of 358 lens candidates, in order to characterizethe candidate lens galaxies and to search for spectral signaturesof high-redshift background galaxies. This approach was previ-ously used to select the SLACS sample (Bolton et al. 2008) andrelies on spectral features captured within the small, 2 (cid:48)(cid:48) diameteraperture fibers. Our examination results in: –
84 spectra of typical LRGs at intermediate redshift withbright continuum, prominent 4000Å break, deep stellar ab-sorption lines, and non- or very faint [OII] λλ – Seven LRG-like spectra with two or more emission and / orabsorption lines falling at a di ff erent and concordant redshift,higher than the LRG redshift, and indicating the presence ofa background galaxy. Two of them are already published asconfirmed strong lens systems: PS1J2343 − + – Eight LRG-like spectra overlaid with a single bright, high-redshift emission line consistent with [OII] λλ . < z < .
50. Althoughthe line widths and resolved double-peaked profiles are thoseexpected for the [OII] doublet, other bright emission lines(H βλ λ λ – Three cases showing clear signatures of star-forming galax-ies at z < .
3, which likely have blue arms misidentified aslensed arcs, one QSO at z = . and only 3 /
104 star-forminggalaxies at lower redshift demonstrates the validity of ourmethod. This result shows that our CNN and visual inspection Although we can not completely rule out that in few cases, theseare actually the central bulges of late-type galaxies covered by BOSS. method predominantly selects the targeted population of galaxy-scale strong lens candidates with lens LRGs, and e ffi ciently dis-tinguishes strong lensing features from usual interlopers (e.g.,spiral arms, tidal features, blue rings). Moreover, the five falsepositives (PS1J2249 + + + + + ≤ . (cid:48)(cid:48) BOSS fibers enclose littlelensed source emission in the θ E (cid:38) . (cid:48)(cid:48) systems targeted by oursearch. Background line flux can only be detected in few fa-vorable cases, like in the presence of counter-images closer tothe lens center. In addition, emission lines in most lensed galax-ies might be too faint for SDSS spectroscopy, and the observ-able spectral band of BOSS excludes H β z > .
9, which rules out multiple line detectionsfor the majority of sources expected at z > + z d = . λ obs > z s = . z s = . z d = . z s = . Z = . (cid:12) cor-rectly fits the overall wavelength range in Fig. 10. Together with
14. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search the red colors of multiple images this adds further evidence thatthe lensed galaxy has indeed ceased star formation recently.The last four CNN candidates showing multiple emissionlines from a background, possibly lensed galaxy do not ex-hibit such clear configurations in Pan-STARRS and Legacy gri images. Their BOSS spectra reveal simultaneous detections of[OII] λλ βλ λ λ − z LRG = . z s = . + z LRG = . z s = . + z LRG = . z s = . z LRG , z s , and σ ∗ LRG suggest plau-sible Einstein radii ∼ (cid:48)(cid:48) for singular isothermal sphere pro-files, high-resolution follow-up imaging is needed to ascertaintheir configuration. These data will determine whether the back-ground, spectroscopically-confirmed galaxies are those visu-ally identified with Pan-STARRS and indeed multiply-imaged,or whether they are out of the strong lensing regime. Lastly,PS1J1724 + z LRG = . z s = . Our visual inspection stage implies that resulting lens candi-dates have morphologies and configurations easily detectableby the human eye. Figure 9 confirms that higher gradesare given to systems with extended arcs, clear counter im-ages and often compact lens light profiles. Postage stampsmake it clear that our most reliable systems have relativelybrighter and distorted arcs (as plausible strongly lensed back-ground galaxies) in gri bands which will facilitate their fu-ture follow-up. Interestingly, besides the numerous blue arcs(e.g., PS1J1647 + − − + − ∼
3. Other biases on the relative positionsof lenses and sources might also arise. For instance, although ourtests suggest that systems with θ E / R e ff < ff erences in p CNN with respectto θ E / R e ff > θ E / R e ff distribution. Quantifying this e ff ectis nevertheless an arduous task due to uncertainties in reliablymeasuring R e ff from Pan-STARRS imaging. As previously men-tioned, systems with θ E di ff ering from simulated lenses in ourtraining set are likely all discarded.Our 358 candidates have lens redshifts in the range z d ∼ . z spec from SDSS, and thosewith more uncertain SDSS z phot due to lens / source blend-ing. It closely matches the lens redshift ranges in other sam-ples selected through multiband, ground-based imaging such asCASSOWARY (Stark et al. 2013) and SL2S (Sonnenfeld et al.2013). Interestingly, Pan-STARRS candidates with SDSS stel-lar velocity dispersion have σ ∗ mean ∼
275 km s − . This is signif-icantly lower than σ ∗ mean ∼
310 km s − in the set of LRGs for simulations (Sect. 4.4), indicating that lens LRGs in our candi-date set are less massive on average.To understand implications of such di ff erences in σ ∗ we fol-lowed Petrillo et al. (2017) and, for each lens candidate with z d and σ ∗ available from BOSS, we calculated Einstein radii for asingular isothermal sphere profile and di ff erent source redshiftsspanning the most plausible z s = . σ SIS ∼ σ ∗ . In the vast majority of cases (86%), Einstein radiipredicted from the lens galaxy dynamics match the approximate θ E ∼ (cid:48)(cid:48) of PS1 candidates from a basic inspection of 3-colorimages in Fig. 9. This range is also expected from constructionof the training set and our CNN selection function. Predictedand observed image configurations usually match for z s >
1. Thisqualitative test argues in favour of plausible lens configurationsand few contaminants despite the moderate σ ∗ from BOSS.In some cases, the large Einstein radii (cid:38) . (cid:48)(cid:48) are likely tobe partly attributed to a contribution from the environment ofthe primary lens galaxy, probably from a smooth dark-matterhalo associated with an extended group- or cluster-scale struc-ture (Oguri 2006). As a matter of fact, obtaining a few stronglens candidates with θ E in the same range as our lens simula-tions but with lower SDSS stellar velocity dispersions provesthat, if they are genuine strong lenses, they host moderatelymassive lens LRGs with significant external shear and magni-fication contributions from the environment. We verified thatsome of our Pan-STARRS candidates are indeed matching theposition of the central brightest cluster galaxy of SDSS clus-ters (e.g., PS1J0907 + + + + + + + + + ff et al. 2016). Thisenvironmental e ff ect is well illustrated by the case ofPS1J0454 − + ∼ (cid:48)(cid:48) ex-tension lie behind the core of Abell 2422 galaxy cluster (Abellet al. 1989). Lastly, PS1J1142 + + z = . + ff set by 1.3 (cid:48) from thebrightest cluster galaxy.The Pan-STARRS gri photometry of our 358 candidates iscomparable to that of our lens simulations. Observed ( g − i ) col-ors of the deflector central regions within 2 (cid:48)(cid:48) diameter aperturesrange between 1.3 and 2.7 mag and broadly follow the colordistribution of mocks and p cat > . g − i ) colors from blended deflectorand source emissions are ∼ p CNN > . p CNN > .
15. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search
We did not find significant correlations between averagegrades and photometric properties. Although our high and lower-confidence candidates have comparable lens brightness, thosewith higher grades have slightly redder lenses by ∼ g − i ) on average. Mean lens redshifts and velocity dispersionsare also increasing in the top half of the candidate set, by 0.05and 50 km s − , respectively. Our study demonstrates that a two-step approach combiningcatalog-level and image-level classifications, as already appliedto lensed quasars searches (Agnello et al. 2015), is e ffi cient in se-lecting galaxy-scale strong lenses from very wide-field surveyswithout imposing restrictive cuts on the input galaxy catalog.Using three bands and multi-aperture photometry, our catalog-level neural network conveniently reduces the sample size forCNN classification to the order of 1 million sources which, inthe case of Pan-STARRS, allows us to download all gri cutoutswithin a manageable amount of time of about one week. Thisfirst stage will be particularly beneficial to overcome data vol-ume limitations in the new era of deep, large scale surveys suchas LSST. At the same time, the low false negative rate and goodperformance on known lens systems of our catalog-level net-work applied to Pan-STARRS photometry shows that this stagediscards very few potentially interesting systems. After that, weobtain encouraging results from our CNN combined with visualinspection. The compilation of 330 newly-discovered lens can-didates (after removing the five false-positives identified fromSDSS BOSS spectra) and confirmation of 23 lenses alreadyselected from other, more suitable surveys (e.g., DES, KiDS)demonstrates the e ffi ciency of our approach regardless of thelimited depth and seeing in Pan-STARRS.The strategy presented in this paper is expected to be eas-ily applicable to the deeper LSST gri stacks in order to identifythe large set of new, wide-separation galaxy-scale lenses fromthis survey. The simulations of Collett (2015) suggest that theLSST final stacks in gri bands will yield about 39 000 detectablelens systems, with θ E greater than the seeing FWHM and largeenough to spatially-resolve multiple images, and with total S / Nof lensed arcs >
20 in at least one band. Under these assump-tions, future LSST lenses will mainly cover θ E = (cid:48)(cid:48) , a simi-lar range as our Pan-STARRS candidates, and their systematicidentification will be crucial for several science cases includingthe search for strongly lensed SNe with spatially-resolved mul-tiple images for early-phase SN spectroscopy and cosmography(Suyu et al. 2020). In order to achieve this, we can customizeour simulation pipeline of realistic Pan-STARRS mock lensesto produce LSST mocks and quickly assemble a training set, byusing gri cutouts of the relevant LRG lenses from LSST stacksrather than Pan-STARRS. Extending the search to broader pop-ulations of lens galaxies might also become feasible, in light ofthe ∼ gri LSST stacks.Tests on the CNN performance presented in this paper arelikely valid essentially for Pan-STARRS seeing and depth, whilebetter imaging quality should enable identifications of broader,more complex image patterns and would likely benefit from up-dates on the neural network. The main challenge is the min-imization of false positives which, given the rarity of stronglenses per deg , will largely dominate the sample of CNN can-didates even with FPR down to 1% or less. Potential avenuesfor improvements include adding other convolutional layers orboosting the learning process with residual networks (He et al.2015), as the latter might give better performance for image fea- tures not represented in the training data (for LSST simulations,see Lanusse et al. 2018). A possibility would be to use CNNsoptimized for image outlier detection (e.g., Margalef-Bentabolet al. 2020), either to identify strong lenses as the outlier class,or simply to exclude cutouts with partial coverage, backgroundartifacts or other anomalies before classification. In addition,although Teimoorinia et al. (2020) found little di ff erence withsingle-band HST images, implementing additional classes forthe usual interlopers (spirals, ring galaxies, ...) might help distin-guish signatures of strong lenses in gri bands. Our Pan-STARRSlens search demonstrates that ultimately, most progress shouldcome from the choice of positive and negative examples in thetraining set (see Sect. 5.5). Highly-realistic lens simulations areone of the main ingredients (see Lanusse et al. 2018). For in-stance, as multiple images are strongly blended with lenses inabout half of the mocks produced in Sect. 4.4, one could only in-clude those visually classified as definite lenses from their brightarcs and unambiguous configurations. Fine-tuning the selectionof positive examples in this way could help the network assignhigh scores exclusively to robust candidates that are worth in-cluding for follow-up campaigns.Reducing the false positive rates and making sure that onlyhigh-confidence candidates get p CNN ∼ ffi cientneural networks should provide orders of magnitudes more lenscandidates than Pan-STARRS and their systematic visual in-spection might become unrealistic. Assembling complete stronglens samples will therefore require one to abandon this non-automated, human-classification stage and to search for alterna-tives. For instance, calibrating the CNN scores as probabilities(e.g., Guo et al. 2017) could help quantify purity and complete-ness, and automatically select a subset of candidates for high-resolution imaging or spectroscopic follow-up.
7. Conclusion
In this paper, we present a systematic search for wide-separation,galaxy-scale strong lenses in Pan-STARRS using machine learn-ing classification of gri images from the 3 π survey on the entirenorthern sky. We focused our search on massive LRGs actingas strong lenses and producing Einstein radii ≥ . (cid:48)(cid:48) , and wesimulated a set of highly-realistic mocks by painting lensed arcson Pan-STARRS image cutouts of LRGs with known redshiftand velocity dispersion from SDSS. This strategy ensures thatmocks include background artifacts and field galaxies, and thatthe network becomes invariant under the small scale exposureand FWHM variations on the survey stacks.We computed the gri aperture photometry of mocks to pre-select a conservative catalog of 23.1 million sources, and fol-lowed a two-step approach for classification: (1) a catalog-basedneural network on the source photometry, (2) a CNN trained on gri image cutouts. The catalog-level network assigned p cat > . p CNN > . > .
9, respectively. We visually inspected those with p CNN > . ≥
2. Publicly available BOSS spectroscopy of thelens candidates’ central regions proves that the vast majorityare indeed LRGs at z ∼ .
16. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search a quadruply-imaged red galaxy at z s = .
185 (likely recentlyquenched), behind a lens LRG at z d = . ffi culties a ff ecting θ E < (cid:48)(cid:48) lensed SNe such as iPTF16geu and allows one to measure accu-rate, microlensing-free time delays for cosmography. Such timedelays are also well-suited to trigger timely imaging and spec-troscopic follow-up of a SN reappearance to characterize theSN’s early-phase behaviour within a few days after explosion.In the meantime, validating these Pan-STARRS lens candidatesand deriving strong lensing models will require high-resolutionimaging and / or spectroscopic follow-up.Our CNN exhibits good performance on known lenses de-tected in Pan-STARRS, correctly classifying 14 /
16 systems aslenses and assigning scores p CNN > . /
16. We found thatCNN predictions strongly depend on the construction of thetraining set but little on the network architecture. Our search alsorecovered 23 confirmed or candidate strong lenses in the litera-ture out of the MasterLens catalog. In the near future, the releaseof deep stacks with best seeing conditions in smaller fields aspart of Pan-STARRS DR3 will give the opportunity to extendour search to optimal PS1 data quality. In addition, we expectthat the e ffi cient and automated two-step classification methodpresented in this paper will be applicable to the deep gri imagestacks from LSST with only minor adjustments. Acknowledgements
We would like to thank Matt Auger and Alessandro Sonnenfeldfor sharing python scripts for spectrum visualisation. We alsothank B. Cl´ement, F. Courbin, S. Huber, C.-H. Lee, D. Sluse andA. Yıldırım for useful discussions and feedback about this work.RC, SS and SHS thank the Max Planck Society for supportthrough the Max Planck Research Group for SHS. This projecthas received funding from the European Research Council(ERC) under the European Union’s Horizon 2020 researchand innovation programme (LENSNOVA: grant agreement No771776). This works has also been in part supported by theSwiss National Science Foundation (SNSF) and by the EuropeanResearch Council (ERC) under the European Union’s Horizon2020 research and innovation program (COSMICLENS: grantagreement No 787886).The Pan-STARRS1 Surveys (PS1) and the PS1 public sci-ence archive have been made possible through contributions bythe Institute for Astronomy, the University of Hawaii, the Pan-STARRS Project O ffi ce, the Max-Planck Society and its par-ticipating institutes, the Max Planck Institute for Astronomy,Heidelberg and the Max Planck Institute for ExtraterrestrialPhysics, Garching, The Johns Hopkins University, DurhamUniversity, the University of Edinburgh, the Queen’s UniversityBelfast, the Harvard-Smithsonian Center for Astrophysics,the Las Cumbres Observatory Global Telescope NetworkIncorporated, the National Central University of Taiwan, theSpace Telescope Science Institute, the National Aeronautics andSpace Administration under Grant No. NNX08AR22G issuedthrough the Planetary Science Division of the NASA ScienceMission Directorate, the National Science Foundation GrantNo. AST-1238877, the University of Maryland, Eotvos LorandUniversity (ELTE), the Los Alamos National Laboratory, and theGordon and Betty Moore Foundation.Based in part on data collected at the Subaru Telescopeand retrieved from the HSC data archive system, which is operated by Subaru Telescope and Astronomy Data Centerat National Astronomical Observatory of Japan. The HyperSuprime-Cam (HSC) collaboration includes the astronomicalcommunities of Japan and Taiwan, and Princeton University.The HSC instrumentation and software were developed by theNational Astronomical Observatory of Japan (NAOJ), the KavliInstitute for the Physics and Mathematics of the Universe (KavliIPMU), the University of Tokyo, the High Energy AcceleratorResearch Organization (KEK), the Academia Sinica Institute forAstronomy and Astrophysics in Taiwan (ASIAA), and PrincetonUniversity. Funding was contributed by the FIRST programfrom Japanese Cabinet O ffi ce, the Ministry of Education,Culture, Sports, Science and Technology (MEXT), the JapanSociety for the Promotion of Science (JSPS), Japan Scienceand Technology Agency (JST), the Toray Science Foundation,NAOJ, Kavli IPMU, KEK, ASIAA, and Princeton University.This paper makes use of software developed for the RubinObservatory Legacy Survey of Space and Time. We thank theLSST Project for making their code available as free software athttp: // dm.lsst.org. References
Abell, G. O., Corwin, Harold G., J., & Olowin, R. P. 1989, ApJS, 70, 1Agnello, A., Kelly, B. C., Treu, T., & Marshall, P. J. 2015, MNRAS, 448, 1446Aihara, H., AlSayyad, Y., Ando, M., et al. 2019, PASJ, 106Aihara, H., Armstrong, R., Bickerton, S., et al. 2018, PASJ, 70, S8Auger, M. W., Budzynski, J. M., Belokurov, V., Koposov, S. E., & McCarthy,I. G. 2013, MNRAS, 436, 503Auger, M. W., Treu, T., Bolton, A. S., et al. 2009, ApJ, 705, 1099Avestruz, C., Li, N., Zhu, H., et al. 2019, ApJ, 877, 58Barnab`e, M., Czoske, O., Koopmans, L. V. E., Treu, T., & Bolton, A. S. 2011,MNRAS, 415, 2215Barnab`e, M., Spiniello, C., Koopmans, L. V. E., et al. 2013, MNRAS, 436, 253Barrena, R., Girardi, M., Boschin, W., De Grandi, S., & Rossetti, M. 2014,MNRAS, 442, 2216Bautista, J. E., Vargas-Maga˜na, M., Dawson, K. S., et al. 2018, ApJ, 863, 110Berghea, C. T., Nelson, G. J., Rusu, C. E., Keeton, C. R., & Dudik, R. P. 2017,ApJ, 844, 90Bertin, E. & Arnouts, S. 1996, A&AS, 117, 393Bertin, E., Mellier, Y., Radovich, M., et al. 2002, Astronomical Society ofthe Pacific Conference Series, Vol. 281, The TERAPIX Pipeline, ed. D. A.Bohlender, D. Durand, & T. H. Handley, 228Bolton, A. S., Burles, S., Koopmans, L. V. E., et al. 2008, ApJ, 682, 964Bonvin, V., Courbin, F., Suyu, S. H., et al. 2017, MNRAS, 465, 4914Bonvin, V., Tihhonova, O., Millon, M., et al. 2019, A&A, 621, A55Bruzual, G. & Charlot, S. 2003, MNRAS, 344, 1000Buckley-Geer, E. J., Lin, H., Rusu, C., et al. 2020, arXiv e-prints,arXiv:2003.12117Ca˜nameras, R., Nesvadba, N., Kneissl, R., et al. 2017a, A&A, 604, A117Ca˜nameras, R., Nesvadba, N. P. H., Kneissl, R., et al. 2017b, A&A, 600, L3Cabanac, R. A., Alard, C., Dantel-Fort, M., et al. 2007, A&A, 461, 813Cava, A., Schaerer, D., Richard, J., et al. 2018, Nature Astronomy, 2, 76Chambers, K. C., Magnier, E. A., Metcalfe, N., et al. 2016, arXiv e-prints,arXiv:1612.05560Chan, J. H. H., Suyu, S. H., Sonnenfeld, A., et al. 2019, arXiv e-prints,arXiv:1911.02587Cheng, T.-Y., Li, N., Conselice, C. J., et al. 2019, arXiv e-prints,arXiv:1911.04320Collett, T. E. 2015, ApJ, 811, 20Comparat, J., Richard, J., Kneib, J.-P., et al. 2015, A&A, 575, A40Courbin, F., Bonvin, V., Buckley-Geer, E., et al. 2018, A&A, 609, A71Danielson, A. L. R., Swinbank, A. M., Smail, I., et al. 2011, MNRAS, 410, 1687Davies, A., Serjeant, S., & Bromley, J. M. 2019, MNRAS, 487, 5263Dawson, K. S., Schlegel, D. J., Ahn, C. P., et al. 2013, AJ, 145, 10de Jong, J. T. A., Verdoes Kleijn, G. A., Kuijken, K. H., & Valentijn, E. A. 2013,Experimental Astronomy, 35, 25Dey, A., Schlegel, D. J., Lang, D., et al. 2019, AJ, 157, 168Dhawan, S., Johansson, J., Goobar, A., et al. 2020, MNRAS, 491, 2639Diehl, H. T., Buckley-Geer, E. J., Lindgren, K. A., et al. 2017, ApJS, 232, 15Dieleman, S., Willett, K. W., & Dambre, J. 2015, MNRAS, 450, 1441Eisenstein, D. J., Annis, J., Gunn, J. E., et al. 2001, AJ, 122, 2267Falco, E. E., Gorenstein, M. V., & Shapiro, I. I. 1985, ApJ, 289, L1
17. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search
Farrow, D. J., Cole, S., Metcalfe, N., et al. 2014, MNRAS, 437, 748Faure, C., Kneib, J.-P., Covone, G., et al. 2008, ApJS, 176, 19Flewelling, H. A., Magnier, E. A., Chambers, K. C., et al. 2016, arXiv e-prints,arXiv:1612.05243Freedman, W. L., Madore, B. F., Hatt, D., et al. 2019, ApJ, 882, 34Freedman, W. L., Madore, B. F., Hoyt, T., et al. 2020, ApJ, 891, 57Gavazzi, R., Marshall, P. J., Treu, T., & Sonnenfeld, A. 2014, ApJ, 785, 144Gavazzi, R., Treu, T., Rhodes, J. D., et al. 2007, ApJ, 667, 176Goldstein, D. A. & Nugent, P. E. 2017, ApJ, 834, L5Goldstein, D. A., Nugent, P. E., Kasen, D. N., & Collett, T. E. 2018, ApJ, 855,22Goobar, A., Amanullah, R., Kulkarni, S. R., et al. 2017, Science, 356, 291Gri ffi th, R. L., Cooper, M. C., Newman, J. A., et al. 2012, ApJS, 200, 9Grillo, C., Gobat, R., Lombardi, M., & Rosati, P. 2009, A&A, 501, 461Grillo, C., Rosati, P., Suyu, S. H., et al. 2018, ApJ, 860, 94Grillo, C., Rosati, P., Suyu, S. H., et al. 2020, arXiv e-prints, arXiv:2001.02232Grogin, N. A., Kocevski, D. D., Faber, S. M., et al. 2011, ApJS, 197, 35Guo, C., Pleiss, G., Sun, Y., & Weinberger, K. Q. 2017, in Proceedings ofMachine Learning Research, Vol. 70, Proceedings of the 34th InternationalConference on Machine Learning, ed. D. Precup & Y. W. Teh (InternationalConvention Centre, Sydney, Australia: PMLR), 1321–1330Hakobyan, A. A., Adibekyan, V. Z., Aramyan, L. S., et al. 2012, A&A, 544, A81Halkola, A., Seitz, S., & Pannella, M. 2007, ApJ, 656, 739Hasinger, G., Capak, P., Salvato, M., et al. 2018, ApJ, 858, 77He, K., Zhang, X., Ren, S., & Sun, J. 2015, arXiv e-prints, arXiv:1512.03385Hezaveh, Y. D., Dalal, N., Marrone, D. P., et al. 2016, ApJ, 823, 37Huang, X., Domingo, M., Pilon, A., et al. 2019, arXiv e-prints,arXiv:1906.00970Huber, S., Suyu, S. H., Noebauer, U. M., et al. 2019, A&A, 631, A161Io ff e, S. & Szegedy, C. 2015, in Proceedings of the 32nd InternationalConference on International Conference on Machine Learning - Volume 37,ICML15 (JMLR.org), 448456Ivezi´c, ˇZ., Kahn, S. M., Tyson, J. A., et al. 2019, ApJ, 873, 111Jacobs, C., Collett, T., Glazebrook, K., et al. 2019a, ApJS, 243, 17Jacobs, C., Collett, T., Glazebrook, K., et al. 2019b, MNRAS, 484, 5330Jacobs, C., Glazebrook, K., Collett, T., More, A., & McCarthy, C. 2017,MNRAS, 471, 167Jaelani, A. T., More, A., Oguri, M., et al. 2020, arXiv e-prints, arXiv:2002.01611Kelly, P. L., Rodney, S. A., Treu, T., et al. 2015, Science, 347, 1123Koekemoer, A. M., Aussel, H., Calzetti, D., et al. 2007, ApJS, 172, 196Koekemoer, A. M., Faber, S. M., Ferguson, H. C., et al. 2011, ApJS, 197, 36Koopmans, L. V. E., Bolton, A., Treu, T., et al. 2009, ApJ, 703, L51Laigle, C., McCracken, H. J., Ilbert, O., et al. 2016, ApJS, 224, 24Lanusse, F., Ma, Q., Li, N., et al. 2018, MNRAS, 473, 3895Lasker, B. M., Lattanzi, M. G., McLean, B. J., et al. 2008, AJ, 136, 735Le F`evre, O., Tasca, L. A. M., Cassata, P., et al. 2015, A&A, 576, A79Leauthaud, A., Massey, R., Kneib, J.-P., et al. 2007, ApJS, 172, 219LeCun, Y., Bengio, Y., & Hinton, G. 2015, Nature, 521, 436LeCun, Y., Bottou, L., Bengio, Y., & Ha ff ner, P. 1998, in Proceedings of theIEEE, 2278–2324Lemon, C. A., Auger, M. W., & McMahon, R. G. 2019, MNRAS, 483, 4242Li, R., Napolitano, N. R., Tortora, C., et al. 2020, arXiv e-prints,arXiv:2004.02715Lilly, S. J., Le F`evre, O., Renzini, A., et al. 2007, ApJS, 172, 70Lintott, C., Schawinski, K., Bamford, S., et al. 2011, MNRAS, 410, 166Lintott, C. J., Schawinski, K., Slosar, A., et al. 2008, MNRAS, 389, 1179M. Green, G. 2018, The Journal of Open Source Software, 3, 695Magnier, E. A., Sweeney, W. E., Chambers, K. C., et al. 2016, arXiv e-prints,arXiv:1612.05244Margalef-Bentabol, B., Huertas-Company, M., Charnock, T., et al. 2020, arXive-prints, arXiv:2003.08263Marshall, P. J., Hogg, D. W., Moustakas, L. A., et al. 2009, ApJ, 694, 924Marshall, P. J., Verma, A., More, A., et al. 2016, MNRAS, 455, 1171Masci, F. J., Laher, R. R., Rusholme, B., et al. 2019, PASP, 131, 018003Metcalf, R. B., Meneghetti, M., Avestruz, C., et al. 2019, A&A, 625, A119Millon, M., Galan, A., Courbin, F., et al. 2019, arXiv e-prints, arXiv:1912.08027More, A., Suyu, S. H., Oguri, M., More, S., & Lee, C.-H. 2017, ApJ, 835, L25More, A., Verma, A., Marshall, P. J., et al. 2016, MNRAS, 455, 1191Nair, V. & Hinton, G. E. 2010, in ICML, ed. J. Frnkranz & T. Joachims(Omnipress), 807–814Oguri, M. 2006, MNRAS, 367, 1241Oguri, M. & Marshall, P. J. 2010, MNRAS, 405, 2579Petrillo, C. E., Tortora, C., Chatterjee, S., et al. 2017, MNRAS, 472, 1129Petrillo, C. E., Tortora, C., Vernardos, G., et al. 2019, MNRASPlanck Collaboration. 2018, Planck 2018 results. VI. Cosmological parametersPlanck Collaboration XIII. 2016, A&A, 594, A13Prakash, A., Licquia, T. C., Newman, J. A., et al. 2016, ApJS, 224, 34 Ranzato, M. A., Boureau, Y.-L., & LeCun, Y. 2007, in Proceedings of the 20thInternational Conference on Neural Information Processing Systems, NIPS07(Red Hook, NY, USA: Curran Associates Inc.), 11851192Riess, A. G., Casertano, S., Yuan, W., Macri, L. M., & Scolnic, D. 2019, ApJ,876, 85Ritondale, E., Vegetti, S., Despali, G., et al. 2019, MNRAS, 485, 2179Rumelhart, D. E., Hinton, G. E., & Williams, R. J. 1986, Nature, 323, 533Russakovsky, O., Deng, J., Su, H., et al. 2015, International Journal of ComputerVision (IJCV), 115, 211Rusu, C. E., Berghea, C. T., Fassnacht, C. D., et al. 2019, MNRAS, 486, 4987Ryko ff , E. S., Rozo, E., Hollowood, D., et al. 2016, ApJS, 224, 1Schaefer, C., Geiger, M., Kuntzer, T., & Kneib, J. P. 2018, A&A, 611, A2Schirmer, M., Suyu, S., Schrabback, T., et al. 2010, A&A, 514, A60Schlafly, E. F. & Finkbeiner, D. P. 2011, ApJ, 737, 103Schlegel, D., White, M., & Eisenstein, D. 2009, in astro2010: The Astronomyand Astrophysics Decadal Survey, Vol. 2010, 314Schlegel, D. J., Finkbeiner, D. P., & Davis, M. 1998, ApJ, 500, 525Schuldt, S., Chiriv`ı, G., Suyu, S. H., et al. 2019, A&A, 631, A40Scoville, N., Abraham, R. G., Aussel, H., et al. 2007, ApJS, 172, 38Secrest, N. J., Dudik, R. P., Dorland, B. N., et al. 2015, ApJS, 221, 12Shajib, A. J., Birrer, S., Treu, T., et al. 2020, MNRASShu, Y., Bolton, A. S., Brownstein, J. R., et al. 2015, ApJ, 803, 71Silverman, J. D., Kashino, D., Sanders, D., et al. 2015, ApJS, 220, 12Simmons, B. D., Lintott, C., Willett, K. W., et al. 2017, MNRAS, 464, 4420Sonnenfeld, A., Chan, J. H. H., Shu, Y., et al. 2018, PASJ, 70, S29Sonnenfeld, A., Gavazzi, R., Suyu, S. H., Treu, T., & Marshall, P. J. 2013, ApJ,777, 97Sonnenfeld, A., Jaelani, A. T., Chan, J., et al. 2019, A&A, 630, A71Sonnenfeld, A., Treu, T., Marshall, P. J., et al. 2015, ApJ, 800, 94Sonnenfeld, A., Verma, A., More, A., et al. 2020, arXiv e-prints,arXiv:2004.00634Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., & Salakhutdinov, R.2014, Journal of Machine Learning Research, 15, 1929Stark, D. P., Auger, M., Belokurov, V., et al. 2013, MNRAS, 436, 1040Suyu, S. H., Bonvin, V., Courbin, F., et al. 2017, MNRAS, 468, 2590Suyu, S. H. & Halkola, A. 2010, A&A, 524, A94Suyu, S. H., Hensel, S. W., McKean, J. P., et al. 2012, ApJ, 750, 10Suyu, S. H., Huber, S., Ca˜nameras, R., et al. 2020, arXiv e-prints,arXiv:2002.08378Szabo, T., Pierpaoli, E., Dong, F., Pipino, A., & Gunn, J. 2011, ApJ, 736, 21Tanaka, M., Hasinger, G., Silverman, J. D., et al. 2017, arXiv e-prints,arXiv:1706.00566Tasca, L. A. M., Le F`evre, O., Ribeiro, B., et al. 2017, A&A, 600, A110Teimoorinia, H., Toyonaga, R. D., Fabbro, S., & Bottrell, C. 2020, PASP, 132,044501Tonry, J. & Onaka, P. 2009, in Advanced Maui Optical and Space SurveillanceTechnologies Conference, E40Treu, T. & Koopmans, L. V. E. 2002, ApJ, 575, 87Turner, E. L., Ostriker, J. P., & Gott, J. R., I. 1984, ApJ, 284, 1Tutukov, A. V. & Yungelson, L. R. 1981, Nauchnye Informatsii, 49, 3Vegetti, S., Lagattuta, D. J., McKean, J. P., et al. 2012, Nature, 481, 341Wang, L., Shu, Y., Li, R., et al. 2017, MNRAS, 468, 3757Waters, C. Z., Magnier, E. A., Price, P. A., et al. 2016, arXiv e-prints,arXiv:1612.05245Whelan, J. & Iben, Icko, J. 1973, ApJ, 186, 1007Willett, K. W., Galloway, M. A., Bamford, S. P., et al. 2017, MNRAS, 464, 4176Willett, K. W., Lintott, C. J., Bamford, S. P., et al. 2013, MNRAS, 435, 2835Wojtak, R., Hjorth, J., & Gall, C. 2019, MNRAS, 487, 3342Wong, K. C., Sonnenfeld, A., Chan, J. H. H., et al. 2018, ApJ, 867, 107Wong, K. C., Suyu, S. H., Chen, G. C. F., et al. 2019, arXiv e-prints,arXiv:1907.04869Yahalomi, D. A., Schechter, P. L., & Wambsganss, J. 2017, arXiv e-prints,arXiv:1711.07919 Appendix A: Complete list of candidates Max-Planck-Institut f¨ur Astrophysik, Karl-Schwarzschild-Str. 1,85748 Garching, Germany e-mail: [email protected] Physik Department, Technische Universit¨at M¨unchen, James-FranckStr. 1, 85741 Garching, Germany Institute of Astronomy and Astrophysics, Academia Sinica, 11F ofASMAB, No.1, Section 4, Roosevelt Road, Taipei 10617, Taiwan18. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens searchName RA Dec p CNN
Grade g Kron r Kron i Kron g aper r aper i aper Redshift NotesPS1J0454 − − − − − − + + + + − − ( ∗ ) (h)PS1J2338 − − + + + + + + + + ( ∗ ) PS1J2144 + + + + + + + + + + + + ( ∗ ) PS1J1650 + + ( ∗ ) PS1J1603 + + ( ∗ ) PS1J1527 + + + + ( ∗ ) (f)PS1J1420 + + + + ( ∗ ) PS1J1417 + + ( ∗ ) PS1J1410 + + ( ∗ ) PS1J1247 + + ( ∗ ) PS1J1210 + + + + ( ∗ ) PS1J1031 + + + + ( ∗ ) PS1J1007 − − − − + + ( ∗ ) PS1J0913 + + ( ∗ ) PS1J0909 + + ( ∗ ) PS1J0410 − − + + − − − − − − + + − − ( ∗ ) PS1J2345 − − ( ∗ ) PS1J2327 + + ( ∗ ) PS1J2319 + + ( ∗ ) PS1J2315 + + + + ( ∗ ) PS1J2307 − − − − + + ( ∗ ) PS1J2241 − − + + + + + + + + − − + + + + − − ( ∗ ) (a)PS1J2115 + + − − + + + + + + ( ∗ ) (k)PS1J1610 + + Table A.1.
Final list of galaxy-scale strong lens candidates with lens LRGs from our systematic search in Pan-STARRS. See Table 1 caption. 19. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens searchName RA Dec p CNN
Grade g Kron r Kron i Kron g aper r aper i aper Redshift NotesPS1J1609 − − + + + + ( ∗ ) PS1J1530 + + ( ∗ ) PS1J1444 − − ( ∗ ) PS1J1437 + + + + + + ( ∗ ) PS1J1414 + + ( ∗ ) PS1J1412 + + ( ∗ ) PS1J1356 + + + + ( ∗ ) PS1J1319 − − + + ( ∗ ) PS1J1233 + + ( ∗ ) PS1J1209 + + ( ∗ ) PS1J1152 − − + + ( ∗ ) PS1J1022 − − ( ∗ ) (f)PS1J0948 + + + + + + − − + + + + ( ∗ ) PS1J0755 + + ( ∗ ) PS1J0740 + + + + + + + + − − − − − − − − − − − − − − − − − − − − − − − − − − − − + + + + − − + + − − + + ( ∗ ) PS1J0024 − − + + + + − − + + ( ∗ ) PS1J2337 + + − − + + − − − − − − − − − − − − − − Table A.2. continued.20. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens searchName RA Dec p CNN
Grade g Kron r Kron i Kron g aper r aper i aper Redshift NotesPS1J2215 − − ( ∗ ) PS1J2213 − − + + ( ∗ ) PS1J2157 − − + + − − − − ( ∗ ) PS1J2128 + + − − + + + + + + − − + + + + + + + + + + + + + + + + + + ( ∗ ) PS1J1724 + + ( ∗ ) PS1J1713 + + + + ( ∗ ) PS1J1657 + + + + + + ( ∗ ) PS1J1643 + + ( ∗ ) PS1J1634 + + + + + + + + + + + + + + ( ∗ ) PS1J1603 + + + + ( ∗ ) PS1J1552 + + ( ∗ ) PS1J1551 + + ( ∗ ) PS1J1548 + + + + − − + + + + + + ( ∗ ) PS1J1534 + + − − − − − − + + ( ∗ ) PS1J1438 + + ( ∗ ) PS1J1436 − − + + − − − − + + ( ∗ ) PS1J1404 + + ( ∗ ) PS1J1403 − − + + − − − − − − − − + + Table A.3. continued. 21. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens searchName RA Dec p CNN
Grade g Kron r Kron i Kron g aper r aper i aper Redshift NotesPS1J1327 + + − − ( ∗ ) PS1J1316 + + + + + + ( ∗ ) PS1J1304 + + ( ∗ ) PS1J1301 + + ( ∗ ) PS1J1259 − − − − + + + + ( ∗ ) PS1J1237 + + + + ( ∗ ) PS1J1215 + + ( ∗ ) PS1J1211 + + + + − − + + ( ∗ ) PS1J1149 + + + + ( ∗ ) PS1J1147 + + ( ∗ ) PS1J1142 + + + + ( ∗ ) PS1J1136 + + ( ∗ ) PS1J1134 + + ( ∗ ) PS1J1132 + + − − + + − − ( ∗ ) PS1J1106 + + ( ∗ ) PS1J1105 + + ( ∗ ) PS1J1104 + + − − − − + + + + ( ∗ ) PS1J1033 + + + + + + + + ( ∗ ) PS1J1001 + + − − ( ∗ ) PS1J0934 − − − − + + ( ∗ ) PS1J0931 + + + + + + + + ( ∗ ) PS1J0911 − − + + + + + + + + ( ∗ ) PS1J0857 + + + + + + ( ∗ ) PS1J0834 + + ( ∗ ) PS1J0831 + + ( ∗ ) PS1J0821 − − − − + + + + + + + + Table A.4. continued.22. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens searchName RA Dec p CNN
Grade g Kron r Kron i Kron g aper r aper i aper Redshift NotesPS1J0718 + + + + + + + + − − − − − − + + − − − − − − − − − − − − − − − − − − − − − − − − − − − − + + − − + + + + − − − − − − − − ( ∗ ) PS1J0227 − − − − + + + + − − + + + + − − − − + + − − − − − − − − − − − − + + ( ∗ ) PS1J0100 − − − − − − + + ( ∗ ) PS1J0044 − − − − + + + + + + − − + + Table A.5. continued. 23. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search
Fig. A.1.
Pan-STARRS 3-color gri images of candidates with grades ≥ p CNN > .
9. See Fig. 9 caption.24. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search
Fig. A.2. continued. 25. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search
Fig. A.3. continued.26. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search
Fig. A.4. continued. Technical University of Munich, Department of Informatics,Boltzmann-Str. 3, 85748 Garching, Germany Institute of Physics, Laboratory of Astrophysics, EcolePolytechnique F´ed´erale de Lausanne (EPFL), Observatoire deSauverny, 1290 Versoix, Switzerland 27. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search
Fig. A.5.
Legacy 3-color images of candidates covered in DR8.28. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search
Fig. A.6. continued. 29. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search
Fig. A.7. continued.30. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search
Fig. A.8. continued. 31. Ca˜nameras et al.: HOLISMOKES – II. Pan-STARRS lens search
Fig. A.9.
PS1 lens candidates with multiple emission lines from a high-redshift background galaxy overlaid on the SDSS BOSS spectrum ofthe foreground LRG. The black, grey and blue lines show the observed spectrum, the 1 σ noise level, and the best-fit SDSS template for theLRG, respectively. The spectrum is zoomed on the spectral features associated with the background line-emitter rather than with the LRG. Top:
PS1J2345 − z LRG = . z s = . Middle:
PS1J1134 + z LRG = . z s = . Bottom:
PS1J1724 + z LRG = . z s = ..