Re-Classification of HST Targets

B. Pirenne

ST-ECF, European Southern Observatory, Karl-Schwarzschild-Str. 2, D-85748 Garching, Germany. Email: bpirenne@eso.org

F. Murtagh

ST-ECF, European Southern Observatory, Karl-Schwarzschild-Str. 2, D-85748 Garching, Germany. Affiliated to Astrophysics Division, Space Science Department, European Space Agency

Abstract:

1. Introduction

For the retrieval of a class of images, for a survey or for further selection from candidate objects, a prior classification of the data can be very useful. No such classification scheme is currently available to the user, in regard to Hubble Space Telescope (HST) data. However textual, coordinate and other information is certainly available. We therefore took this information, and derived an object classification assignment (or set of assignments) for each target and exposure in the HST image database. In a first phase of this work, we have kept to spectral datasets. In order to benefit from previous classification work, we based this initiative on the International Ultraviolet Explorer (IUE) classification scheme. Apart from focusing our efforts, this has the great advantage of bringing together particular semantic aspects of the contents of two different data collections. The system set up can be seen at http://www.eso.org/hst-query-by-class.html

2. Classification of HST Spectra: Three Phases

The two-digit IUE classification scheme was availed of. This was enhanced to cater for various extra-galactic objects which are now of importance, but which had not played a role back in the days when the IUE scheme was defined. Similarly, adjustments to the IUE classification scheme were made in regard to calibration terminology. The enhanced IUE scheme, now used by us, can be seen at the address given in the previous section.

The following processing phases were carried out in sequence.

1. A list of about 20 names or partial names of planets and other solar system objects was used to search for matches in the object name which accompanied the exposure. Clearly, we cannot depend on coordinate information for such objects!

2. Next, the coordinates were used. To facilitate an approximate match with coordinates of objects in the IUE catalog (or a fuzzy join, in database terms), a translation into galactic latitude and longitude (decimal values) was made. The matching was carried out in a square 0.01 degree area. Approximately 26000 IUE objects were available.

3. Approximately 90 terms were selected from a perusal of observing proposal abstracts, or from the IAU Thesaurus. These were object names, or parts of object names, or abbreviated names. For each, an IAU class was associated. These terms were sought for in the observing proposal abstract, or title, or uncontrolled keyword list.

All classes found for a given spectral dataset were retained. Duplicates were deleted.

3. Results

For FOS, 4847 records were used in October 1995. There was an average of 2 classes per FOS dataset. At most, 14 classes were assigned to a FOS dataset. The number of FOS datasets getting at least one class affiliation was 98.3%.

For GHRS, 6603 records were used in October 1995. There was an average of 3.3 classes per GHRS dataset. At most, 10 classes were assigned to a GHRS dataset. The number of GHRS datasets getting at least one class affiliation was 97.9%.

The procedures used consist of a number of scripts which can be re-run periodically (or even more frequently).

4. The Future

The use of datasets relating to spectra, to date, has facilitated the objective of assigning one class, or a small number of classes, to a target. Application to 2-dimensional image datasets will be more complex. Carrying out an object inventory in all images---using, e.g., packages such as MIDAS's Inventory, or IRAF's FOCAS---would require a good deal of parameter tuning. It may be recalled that, just as for the spectral classification described above, the actual objects imaged would be of interest, and not the stated aim of the observing team. This points to the interest of cross-correlating objects with one or more catalogs, just as we have described above with respect to IUE.

To further expedite the object cataloging we must look further. We would need highly compressed versions of an entire image database. We would need refined object detection methods, which clearly distinguish between noise and significant structure. We would need excellent template matching algorithms (perhaps based on gray-level mathematical morphology). Some work in these directions has been described at previous ADASS meetings. The methods needed to treat entire image databases is now beginning.

A particular topic of note, which will be of greater urgency in the case of two-dimensional image datasets, is that of information fusion. A large number of classification assignments may not be optimally useful to the user. Consensus techniques are a well-researched sub-domain of classification research, modular neural networks, and intelligent pattern recognition systems. Such results can be drawn upon to provide compact semantic descriptions of the complex objects for which we seek class assignments.