Assembly and Classification of Spectral Energy Distributions

1
Assembly and Classification of Spectral Energy
Distributions A New VO Web Service

• Hans-Martin Adorf, GAVO, Max-Planck-Institut für
  extraterr. Physik, Garching
• Florian Kerber, ST-ECF, European Southern
  Observatory, Garching
• Gerard Lemson, GAVO, Max-Planck-Institut für
  extraterr. Physik, Garching
• Alberto Micol, ST-ECF, European Southern
  Observatory, Garching
• Roberto Mignani, European Southern Observatory,
  Garching
• Thomas Rauch, Institut für Astronomie und
  Astrophysik, Universität Tübingen
• Wolfgang Voges, GAVO, Max-Planck-Institut für
  extraterr. Physik, Garching

2
Overview

• We report progress on a new Web service for
  automated object classification which comprises
  four major steps
• An input list of sky-positions is used for
  querying multiple distributed catalogues covering
  different wavelength intervals. The sources
  returned are spatially matched using a
  probabilistic method.
• A list of observed spectral energy distributions
  (SEDs) is assembled.
• The theoretical SEDs are prepared using a library
  of model spectra.
• The obsrvational SEDs are submitted to a
  classifier that uses the theoretical SEDs for
  template matching. For each observed SED the
  three best-matching theoretical SEDs are
  identified.
• A science case has been selected for testing the
  capabilities of the Web service described.
• This work has been carried out as a collaboration
  between the AVO (http//www.euro-vo.org) and GAVO
  (http//www.g-vo.org) projects.

3
Scientific Motivation

• Many scientific investigations benefit from a
  multi-spectral (pan-chromatic) view of the
  universe.
• This idea has played a vital role at the very
  beginning of the virtual observatory movement.
• Some areas of interest
• panchromatic mining for quasars a key-stone
  science application of the US-American NVO.
• AGN research start with a list of AGN
  candidates collect all photometric data from
  distributed catalogues covering the full spectral
  range classify the AGN-zoo (type I, II, BL Lac,
  etc.)
• planetary nebulae, isolated neutron stars, brown
  dwarfs, CVs

4
Catalogue Query and Matching

• The catalogue query and matching process is
  itself a three-stage process
• The user uploads an input list of sky-positions
• The user selects the catalogues of interest. For
  each catalogue a deterministic matching service
  provided by CDS/Vizier is invoked that, for each
  object in the input list, carries out a simple
  cone search.
• The result is a set of match lists, one per
  catalogue. Often the matching results are
  ambiguous.
• Finally, the matcher fuses the match-lists into a
  single master list using GAVOs fuzzy matcher
  algorithm.
• The resulting fused master list contains all
  plausible match candidates. Each entry in this
  list contains at most one source from each
  catalogue.

5
Catalogue Selection
6
Match-List before XMatch
7
Fused Master List (after XMatch)
8
Assembly of the Observational SEDs

• The SED-assembly process for the observational
  data takes several steps
• For each match-candidate the photometric
  measurements are collected from the contributing
  catalogues.
• Since a given catalogue may not have a matching
  source, often the photometric measurements are
  null. Even when the catalogue has a matching
  source there may still be no photometric
  measurement in a given passband.
• Next, unit conversions are applied to the
  photometric measurements in order to form a
  spectral energy distribution (SED).
• The resulting (usually incomplete) SEDs make up
  the features which the classifier operates on.

9
Observation Data Preview
10
Preparation of the Theoretical SEDs

• For the subsequent classification stage, the
  theoretical data has to be brought into the
  observational space.
• We have used a grid of stellar model atmosphere
  spectra (Thomas Rauch, http//astro.uni-tuebingen.
  de/rauch/). The theoretical spectra have a much
  higher resolution than the observational broad
  band SEDs the former therefore have been
  downsampled to match the latter.
• In order to match the low spectral resolution of
  the observations, the theoretical flux was
  extracted at the central wavelength for each of
  the 7 wavebands, i.e. Johnson B, V, R, I, H, J,
  K. (In principle one would have to convert the
  theoretical spectra using the proper sensitivity
  curves of the filters.)
• No correction was applied for interstellar
  extinction

11
Library Data Preview
12
Library Data View
13
Supervised Classification

• The list of observational SEDs is submitted to a
  supervised classifier.
• The classifier uses the library of theoretical
  SEDs for template matching.
• In principle any user-supplied library may be
  used we only require that the uploaded
  theoretical SEDs comprise the same features as
  those in the observed SEDs.
• The SED classifier currently uses a simple
  deterministic nearest neighbour (NN) algorithm
  which uses the Euclidean distance in feature
  space. For each observed SED the NN-classifier
  identifies the three best-matching theoretical
  SEDs.
• User choices the features to use in the
  classification the method for estimating the
  scaling factor the number of best matches to
  report

14
SED Classifier Central
15
Classification View
16
Quick-look Graphics

• For an easy assessment of the results we decided
  to also provide quick-look on-line graphics.
• For each observational SED the chart contains
• the observed SED and
• an overplot of the three best matching
  theoretical SEDs.
• We use the JFreeChart graphics package, wrapped
  in the Cewolf library for use within JavaServer
  Pages (JSPs). Fortunately, only a few lines of
  code are necessary in order to bring up a chart.

17
Quick-look Chart
18
Reporting

• Classification results are reported in a
  classification table.
• the ID of the observational SED,
• the No of (non-null) features contributing to the
  classification, and
• for each of the best three matches
• the ID of the matching theoretical SED,
• the dissimilarity between the observed and the
  matching theoretical SED, and
• a scaling factor (the distance modulus).
• The full complement of pair-wise dissimilarities
  is also reported.
• This table can become very large, since it scales
  with the number of observational SEDs times the
  number of theoretical SEDs.

19
Status

• The SED classifier is implemented in pure Java
• as a standard J2EE Web Application
• We successfully use
• the JavaServer Faces (JSF) technology, which
  offers a server- and a client-side
  state-mechanism,
• We extended it by a custom JSF-tag library for
  table input and output.
• an embedded 100 Java database (HSQLDB) for
  feature selection and reporting, and
• the GAVO table utility package (similar to
  AstroGrids Topcat/STIL package).

20
Conclusions

• A proper handling of missing data (null values)
  is essential for this kind of application.
• Quick-look graphics are helpful to let the user
  assess the classification results.
• We need a statistical classifier to adequately
  handle the photometric uncertainties.
• We need to validate the classifier.
• This is work-in-progress. We are relying on the
  CDS/Vizier matching services, which we extend.

21
Selected References

• Adorf, H.-M. Classification of Low-Resolution
  Stellar Spectra via Template Matching -- A
  Simulation Study. in Workshop "Data Analysis in
  Astronomy II". 1986. Erice, Italy Plenum Press,
  New York, USA.
• Kerber F., Mignani R.P., Guglielmetti F., Wicenec
  A., Galactic Planetary Nebulae and their central
  stars. I. An accurate and homogeneous set of
  coordinates. Astron. Astrophys. 408, 1029 (2003)
• McGlynn, T.A. , A.A. Suchkov, E.L. Winter, R.J.
  Hanisch, R.L. White, F. Ochsenbein, S. Derriere,
  W. Voges, and M.F. Corcoran, Automated
  Classification of ROSAT Sources Using
  Heterogeneous Multi-wavelength Source Catalogs,
  Astrophys. J. (submitted), 2004.
• Padovani, P., Allen, M. G., Rosati, P., Walton,
  N. A. 2004, Discovery of optically faint obscured
  quasars with Virtual Observatory tools, Astronomy
  Astrophys. 424, 545.
• Rauch, T., Grids of Synthetic Stellar Fluxes.
  2004, Thomas Rauch. http//astro.uni-tuebingen.de/
  rauch/.