Automated detection of galaxy-scale gravitational lenses in high-resolution imaging data

Philip J. Marshall, David W. Hogg, Leonidas A. Moustakas, Christopher D. Fassnacht, Marua Brada, Tim Schrabback, Roger D. Blandford

    Research output: Contribution to journalArticle

    Abstract

    We expect direct lens modeling to be the key to successful and meaningful automated strong galaxy-scale gravitational lens detection. We have implemented a lens-modeling "robot" that treats every bright red galaxy (BRG) in a large imaging survey as a potential gravitational lens system. Having optimized a simple model for "typical" galaxy-scale gravitational lenses, we generate four assessments of model quality that are then used in an automated classification. The robot infers from these four data the lens classification parameter H that a human would have assigned; the inference is performed using a probability distribution generated from a human-classified training set of candidates, including realistic simulated lenses and known false positives drawn from the Hubble Space Telescope (HST) Extended Groth Strip (EGS) survey. We compute the expected purity, completeness, and rejection rate, and find that these statistics can be optimized for a particular application by changing the prior probability distribution for H; this is equivalent to defining the robot's "character." Adopting a realistic prior based on expectations for the abundance of lenses, we find that a lens sample may be generated that is 100% pure, but only 20% complete. This shortfall is due primarily to the oversimplicity of the model of both the lens light and mass. With a more optimistic robot, 90% completeness can be achieved while rejecting 90% of the candidate objects. The remaining candidates must be classified by human inspectors. Displaying the images used and produced by the robot on a custom "one-click" web interface, we are able to inspect and classify lens candidates at a rate of a few seconds per system, suggesting that a future 1000 deg2 imaging survey containing 107 BRGs, and some 10 4 lenses, could be successfully, and reproducibly, searched in a modest amount of time. We have verified our projected survey statistics, albeit at low significance, using the HST EGS data, discovering four new lens candidates in the process.

    Original languageEnglish (US)
    Pages (from-to)924-942
    Number of pages19
    JournalAstrophysical Journal
    Volume694
    Issue number2
    DOIs
    StatePublished - 2009

    Fingerprint

    gravitational lenses
    lenses
    galaxies
    high resolution
    robots
    modeling
    completeness
    Hubble Space Telescope
    strip
    detection
    statistics
    inference
    rejection
    purity
    education
    rate
    distribution

    Keywords

    • Galaxies: elliptical and lenticular, cD
    • Gravitational lensing
    • Methods: data analysis
    • Methods: statistical
    • Surveys
    • Techniques: miscellaneous

    ASJC Scopus subject areas

    • Space and Planetary Science
    • Astronomy and Astrophysics

    Cite this

    Marshall, P. J., Hogg, D. W., Moustakas, L. A., Fassnacht, C. D., Brada, M., Schrabback, T., & Blandford, R. D. (2009). Automated detection of galaxy-scale gravitational lenses in high-resolution imaging data. Astrophysical Journal, 694(2), 924-942. https://doi.org/10.1088/0004-637X/694/2/924

    Automated detection of galaxy-scale gravitational lenses in high-resolution imaging data. / Marshall, Philip J.; Hogg, David W.; Moustakas, Leonidas A.; Fassnacht, Christopher D.; Brada, Marua; Schrabback, Tim; Blandford, Roger D.

    In: Astrophysical Journal, Vol. 694, No. 2, 2009, p. 924-942.

    Research output: Contribution to journalArticle

    Marshall, PJ, Hogg, DW, Moustakas, LA, Fassnacht, CD, Brada, M, Schrabback, T & Blandford, RD 2009, 'Automated detection of galaxy-scale gravitational lenses in high-resolution imaging data', Astrophysical Journal, vol. 694, no. 2, pp. 924-942. https://doi.org/10.1088/0004-637X/694/2/924
    Marshall PJ, Hogg DW, Moustakas LA, Fassnacht CD, Brada M, Schrabback T et al. Automated detection of galaxy-scale gravitational lenses in high-resolution imaging data. Astrophysical Journal. 2009;694(2):924-942. https://doi.org/10.1088/0004-637X/694/2/924
    Marshall, Philip J. ; Hogg, David W. ; Moustakas, Leonidas A. ; Fassnacht, Christopher D. ; Brada, Marua ; Schrabback, Tim ; Blandford, Roger D. / Automated detection of galaxy-scale gravitational lenses in high-resolution imaging data. In: Astrophysical Journal. 2009 ; Vol. 694, No. 2. pp. 924-942.
    @article{7f65f1391ba04067836ceb2a85017aa5,
    title = "Automated detection of galaxy-scale gravitational lenses in high-resolution imaging data",
    abstract = "We expect direct lens modeling to be the key to successful and meaningful automated strong galaxy-scale gravitational lens detection. We have implemented a lens-modeling {"}robot{"} that treats every bright red galaxy (BRG) in a large imaging survey as a potential gravitational lens system. Having optimized a simple model for {"}typical{"} galaxy-scale gravitational lenses, we generate four assessments of model quality that are then used in an automated classification. The robot infers from these four data the lens classification parameter H that a human would have assigned; the inference is performed using a probability distribution generated from a human-classified training set of candidates, including realistic simulated lenses and known false positives drawn from the Hubble Space Telescope (HST) Extended Groth Strip (EGS) survey. We compute the expected purity, completeness, and rejection rate, and find that these statistics can be optimized for a particular application by changing the prior probability distribution for H; this is equivalent to defining the robot's {"}character.{"} Adopting a realistic prior based on expectations for the abundance of lenses, we find that a lens sample may be generated that is 100{\%} pure, but only 20{\%} complete. This shortfall is due primarily to the oversimplicity of the model of both the lens light and mass. With a more optimistic robot, 90{\%} completeness can be achieved while rejecting 90{\%} of the candidate objects. The remaining candidates must be classified by human inspectors. Displaying the images used and produced by the robot on a custom {"}one-click{"} web interface, we are able to inspect and classify lens candidates at a rate of a few seconds per system, suggesting that a future 1000 deg2 imaging survey containing 107 BRGs, and some 10 4 lenses, could be successfully, and reproducibly, searched in a modest amount of time. We have verified our projected survey statistics, albeit at low significance, using the HST EGS data, discovering four new lens candidates in the process.",
    keywords = "Galaxies: elliptical and lenticular, cD, Gravitational lensing, Methods: data analysis, Methods: statistical, Surveys, Techniques: miscellaneous",
    author = "Marshall, {Philip J.} and Hogg, {David W.} and Moustakas, {Leonidas A.} and Fassnacht, {Christopher D.} and Marua Brada and Tim Schrabback and Blandford, {Roger D.}",
    year = "2009",
    doi = "10.1088/0004-637X/694/2/924",
    language = "English (US)",
    volume = "694",
    pages = "924--942",
    journal = "Astrophysical Journal",
    issn = "0004-637X",
    publisher = "IOP Publishing Ltd.",
    number = "2",

    }

    TY - JOUR

    T1 - Automated detection of galaxy-scale gravitational lenses in high-resolution imaging data

    AU - Marshall, Philip J.

    AU - Hogg, David W.

    AU - Moustakas, Leonidas A.

    AU - Fassnacht, Christopher D.

    AU - Brada, Marua

    AU - Schrabback, Tim

    AU - Blandford, Roger D.

    PY - 2009

    Y1 - 2009

    N2 - We expect direct lens modeling to be the key to successful and meaningful automated strong galaxy-scale gravitational lens detection. We have implemented a lens-modeling "robot" that treats every bright red galaxy (BRG) in a large imaging survey as a potential gravitational lens system. Having optimized a simple model for "typical" galaxy-scale gravitational lenses, we generate four assessments of model quality that are then used in an automated classification. The robot infers from these four data the lens classification parameter H that a human would have assigned; the inference is performed using a probability distribution generated from a human-classified training set of candidates, including realistic simulated lenses and known false positives drawn from the Hubble Space Telescope (HST) Extended Groth Strip (EGS) survey. We compute the expected purity, completeness, and rejection rate, and find that these statistics can be optimized for a particular application by changing the prior probability distribution for H; this is equivalent to defining the robot's "character." Adopting a realistic prior based on expectations for the abundance of lenses, we find that a lens sample may be generated that is 100% pure, but only 20% complete. This shortfall is due primarily to the oversimplicity of the model of both the lens light and mass. With a more optimistic robot, 90% completeness can be achieved while rejecting 90% of the candidate objects. The remaining candidates must be classified by human inspectors. Displaying the images used and produced by the robot on a custom "one-click" web interface, we are able to inspect and classify lens candidates at a rate of a few seconds per system, suggesting that a future 1000 deg2 imaging survey containing 107 BRGs, and some 10 4 lenses, could be successfully, and reproducibly, searched in a modest amount of time. We have verified our projected survey statistics, albeit at low significance, using the HST EGS data, discovering four new lens candidates in the process.

    AB - We expect direct lens modeling to be the key to successful and meaningful automated strong galaxy-scale gravitational lens detection. We have implemented a lens-modeling "robot" that treats every bright red galaxy (BRG) in a large imaging survey as a potential gravitational lens system. Having optimized a simple model for "typical" galaxy-scale gravitational lenses, we generate four assessments of model quality that are then used in an automated classification. The robot infers from these four data the lens classification parameter H that a human would have assigned; the inference is performed using a probability distribution generated from a human-classified training set of candidates, including realistic simulated lenses and known false positives drawn from the Hubble Space Telescope (HST) Extended Groth Strip (EGS) survey. We compute the expected purity, completeness, and rejection rate, and find that these statistics can be optimized for a particular application by changing the prior probability distribution for H; this is equivalent to defining the robot's "character." Adopting a realistic prior based on expectations for the abundance of lenses, we find that a lens sample may be generated that is 100% pure, but only 20% complete. This shortfall is due primarily to the oversimplicity of the model of both the lens light and mass. With a more optimistic robot, 90% completeness can be achieved while rejecting 90% of the candidate objects. The remaining candidates must be classified by human inspectors. Displaying the images used and produced by the robot on a custom "one-click" web interface, we are able to inspect and classify lens candidates at a rate of a few seconds per system, suggesting that a future 1000 deg2 imaging survey containing 107 BRGs, and some 10 4 lenses, could be successfully, and reproducibly, searched in a modest amount of time. We have verified our projected survey statistics, albeit at low significance, using the HST EGS data, discovering four new lens candidates in the process.

    KW - Galaxies: elliptical and lenticular, cD

    KW - Gravitational lensing

    KW - Methods: data analysis

    KW - Methods: statistical

    KW - Surveys

    KW - Techniques: miscellaneous

    UR - http://www.scopus.com/inward/record.url?scp=70449643673&partnerID=8YFLogxK

    UR - http://www.scopus.com/inward/citedby.url?scp=70449643673&partnerID=8YFLogxK

    U2 - 10.1088/0004-637X/694/2/924

    DO - 10.1088/0004-637X/694/2/924

    M3 - Article

    VL - 694

    SP - 924

    EP - 942

    JO - Astrophysical Journal

    JF - Astrophysical Journal

    SN - 0004-637X

    IS - 2

    ER -