AUTOMATED STAR GALAXY CLASSIFICATION FOR DIGITIZED POSS-II

We describe the automated object classification method implemented in the Sky Image Cataloging and Analysis Tool (SKICAT) and applied to the Digitized Second Palomar Observatory Sky Survey (DPOSS). This classification technique was designed with two purposes in mind: first, to classify objects in DPOSS to the faintest limits of the data; second, to fully generalize to future classification efforts, including classification of galaxies by morphology and improving the existing DPOSS star/galaxy classifiers once a larger volume of data are in hand. To optimize the identification of stars and galaxies in J and F band DPOSS scans, we determined a set of eight highly informative object attributes. In the eight-dimensional space defined by these attributes, we found like objects to be distributed relatively uniformly within and between plates. To infer the rules for distinguishing objects in this, but possibly any other, high-dimensional parameter space, we utilize a machine learning technique known as decision tree induction. Such induction algorithms are able to determine statistical classification rules simply by training on a set of example objects. We used high quality CCD images to determine accurate classifications for those examples in the training and set too faint for reliable classification by visual inspection of the plate. Our initial results obtained from a set of four DPOSS fields indicate that we achieve 90% completeness and 10% contamination in our galaxy catalogs down to a magnitude limit of ∼19.6m in r and 20.5m in g, within F and J plates, respectively, or an equivalent BJ of nearly 21.0m. This represents a 0.5m-1.0m improvement over results from previous digitized Schmidt plate surveys using comparable plate material. We have also begun applying methods of unsupervised classification to the DPOSS catalogs, allowing the data, rather than the scientist, to suggest the relevant and distinct classes within the sample. Our initial results from these experiments suggest the scientific promise of such machine discovery methods in astronomy. © 1995 American Astronomical Society.