Object classification and physical parametrization with gaia and other large surveys

Object classification and physical parametrization with GAIA and other large surveys

• Coryn A.L. Bailer-Jones

• Max-Planck-Institut für Astronomie, Heidelberg

• [e-mail]

Science with surveys

• Survey characteristics

• large numbers of objects (>106)

• no pre-selection  different types of objects

• (stars, galaxies, quasars, asteroids, etc.)

• several observational ‘dimensions’ (e.g. filters, spectra)

• Goals

• discrete classification of objects (star, galaxy; or stellar types)

• continuous physical parametrization (Teff, logg, [Fe/H], etc.)

• efficient detection of new types of objects

• SDSS, LSST, VST/VISTA, DIVA, GAIA, virtual observatory ...

GAIA Galaxy survey mission

• Composition, formation and evolution of our Galaxy

• High precision astrometry for distances and proper motions (10 as @ V=15  1% distance at 1kpc)

• Observe entire sky down to V=20 @ 0.1–0.5´´ resolution

•  109 stars across all stellar populations

• + 105 quasars, 107 galaxies, 105 SNe, 106 SSOs

• Observe everything in 15 medium and broad band filters

• High resolution spectroscopy (for radial velocities) for V<17

• Comparison to Hipparcos:

• ×10 000 objects, ×100 precision, 11 mags deeper

• ESA mission, “approved” for launch in c. 2011

GAIA satellite and mission

• 8.5m × 2.9m (deployed sun shield)

• 3100 kg (at launch)

• Earth-Sun L2 Lissajous orbit

• Continuously rotating (3hr period), precessing (80 days) and observing

• 5 year mission

• Each object observed c.100 times

• Cost at completion: 570 MEuro

GAIA scientific payload

• High stability SiC structure

• Non-deployable 3-mirror telescopes

• Optical (200-1000nm)

• Two astrometric telescopes:

• 1.7m×0.7m, 0.6°×0.6° FOV

• Spectroscopic telescope:

• 0.75m×0.7m, 1°×4° FOV

GAIA astrometric focal plane

• CCDs clocked in TDI mode

• 60cm × 70 cm, 250 CCDs,

• 2780 pixels × 2150 pixels

• 21.5s crossing time

• Star mappers:

• real-time onboard detection

• (only samples transmitted due to limited telemetry rate)

• Main astrometric field:

• high precision centroiding

• (0.001 pix) from high SNR

• Four broad band filters:

• chromatic correction

GAIA spectroscopic focal plane

• Operates on same principle as astrometric field (independent star mappers)

• Light dispersed in across-scan direction in central part of field:

• ~ 1Å resolution spectroscopy around CaII (850-875nm) for V<17

•  1-10 km/s radial velocities, abundances

• 11 medium band filters for all objects

•  object classification, physical parameters, extinction, absolute fluxes

Classification goals for GAIA

• Classification as star, galaxy, quasar, solar system objects etc.

• Determination of physical parameters of all stars

• - Teff, logg, [Fe/H], [/Fe], CNO, A(), Vrot, Vrad, activity

• Use all data (photometric, spectroscopic, astrometric)

• Combine with parallax to determine stellar:

• - luminosity, radius, (mass, age)

• Must be able to cope with:

• - unresolved binaries (help from astrometry)

• - photometric variability (can exploit: Cepheids, RR Lyrae)

• - redshifted objects

• - extended object (can deal with separately)

Classification/Parametrization Principles

• Partition multidimensional data space to:

• 1. classify objects into known classes

• 2. parametrize objects on continuous physical scales

• Assign classes/parameters in presence of noise

• Multiple 2-dimensional colour-colour diagrams inadequate!

• 1. direct probabilistic methods (Goebel et al. 1989; Christlieb et al. 1998)

• neural networks (Storrie-Lombardi et al. 1992; Odewahn et al. 1993)

• clustering methods

• 2. neural networks (Weaver & Torres-Dodgen 1995,1997; Singh et al. 1998

• Bailer-Jones 1996,2000; Snider et al. 2001)

• MDM (Katz et al. 1998; Elsner et al. 1999; Vansevicius et al. 2002)

• Gaussian Processes (krigging) (Bailer-Jones et al. 1999)

Neural Networks (NNs)

• Functional mapping:

• parameters = f(data; weights)

• Weights determined by training on pre-classified data

•  least squares minimization of

• total classification error

•  global interpolation of data

• Problems:

• local minima

• training data distribution

• missing and censored data

Minimum Distance Methods (MDMs)

• Assign parameters according to nearest template(s) (k-nn, 2 min.)

• Generally interpolate:

• either in data space:  = f(d; w)

• or in parameter space: D = g(; w)

•   new =  which minimizes D

• Local methods

• Problems:

• distance weighting

• number of neighbours (bias/variance)

• simultaneous determination of multiple parameters

• speed? (109 in c. 1 week  1500/s)

•  = astrophysical parameter; d = data

Challenges for large, deep surveys

• General

• interstellar extinction

• photometric variability (pulsating stars, quasars)

• multiple solutions (data degeneracy: noise dependent)

• incorporation of prior information (iterative solutions)

• robust to missing and censored data

• known noise model: uncertainty predictions

• template/training data: real vs. synthetic vs. mix

• Additional for GAIA (and DIVA)

• unresolved binary stars (biases parameters)

• use parallax information and local astrometry/RVs

• Most work to date has been on ‘cleaned’ (i.e. biased) data sets

Summary

• Large, deep surveys produce complex, inhomogeneous, multi-dimensional datasets

• Powerful, robust, automated methods for object classification and physical parametrization are required, but ...

• ... many issues remain to be addressed

• GAIA presents particular challenges:

• photometric, spectroscopic, astrometric and kinematic data

• broad science goals  wide range of objects to be classified

• Discrete vs. continuous, local vs. global methods

• (NNs, MDMs, GPs, clustering methods)

• Existing methods to be extended; new methods to be explored

• New members of GAIA Classification WG always welcome!