When a sky survey is made, the telescope images different parts of the sky at different times, wavebands, etc., creating a mosaic of images. Traditionally, the images are converted to star/galaxy catalogs through source-extraction, and much of the scientific knowledge is mined from these catalogs. Catalogs are federated by cross-matching, and further knowledge mined from the joint catalog. However, the intrinsic limitation of any source extraction process is that object detection relies on some filter, which is usually optimized for a particular source size and/or contrast, and may not be efficient in detecting useful signals which fall outside its parameter range; an example are low surface brightness objects. Thus, images may contain valuable information not present in the source catalogs.
In this paper, we consider a new paradigm for mining knowledge from the images of the sky surveys: by federating the images directly to a uniform pixel space, so that multiple image sets can be stacked, then doing source detection, pattern matching and other data mining in the resulting multi-channel image. If the pixel spaces ( pages) are standardized, then it is much easier to federate data made by different groups in different ways.
Even with an "optimal" source extraction, in some cases even the absence of a statistically significant signal represents a useful information, but in the context of other imagery; an example are high-redshift objects detected with the "dropout" techniques. Image federation is thus useful in the detection of faint sources [Szalay 1998], which may not be detected in a particular (or any given) channel, but may be detected in others; or, a collective of individually marginal detections may have a high significance when considered jointly. We can go fainter in image space because we have more photons from the combined images and because the multiple detections can be used to enhance the reliability of sources at a given threshold.
Astronomers are more and more reaching into the time domain (synoptic), imaging the sky repeatedly for discovery of transient and variable sources. Near-Earth asteroids and Kuiper-belt objects, cataclysmic variables, active galactic nuclei, and quasars are all interesting, as well as lensing events, orphan afterglows, and most exciting would be a rich set of new types of transients. Image federation shows promise for analysis of this kind of data, by forming an average sky, a maximal brightness sky, a sky specialized to non-Gaussian noise sources, and a sky where information from dozens of channels is represented in a single color image. Very faint periodic sources could be found by time autocorrelation of pixel brightness.
A prerequisite for finding differences between sky images is the perfect mutual positioning of sources. We will focus on reconciling images with very different inherent spatial resolution by micro-optimization of astrometry through a fitting process. More generally, we can subtract out ``well-fitted'' objects (e.g., PSF-like), to leave only the faint, complex structure that may contain new knowledge.
Image federation enables robust detection and flux measurement of complex, extended sources over a range of size, wavelength, and time scales. Larger objects in the sky may have both extended structure (requiring image mosaicking) and a much smaller active center, or diffuse structure entirely. We will be able to combine multiple instrument imagery to build a multi-scale, multi-wavelength picture of such extended objects. We will also make statistical studies of extended, complex sources that vary in shape with wavelength.
Part of the atlas definition is the collection of pointing centers of its pages. In Hyperatlas, there are currently two choices for this layout: TM and HV. The former puts pointing centers on lines of constant declination, the latter on the centers of Hierarchical Triangular Mesh (HTM) [Kunzst 1999] triangles. Each layout comes with a parameter: for the TM layout, the parameter means the separation between the declination lines; and for the HV layout it is the level of the HTM hierarchy. In figure 1 is an atlas made from the TM-5 configuration of pointing centers. Each of these layout types is defined by an algorithm or pseudocode (see [Williams 2003]), and also available through redundantly-implemented web services.
The atlas is further specified by the projection type (SIN, TAN, AIT, etc. from the WCS library), and by the nominal scale at the pointing center. We have discretized the scale numbers in powers of two: scale 20 is one arcsecond per pixel, and generally scale is arcseconds per pixel.
For example the atlas TM-5-TAN-20 consists of 1734 pages covering the celestial sphere, with every point of the sphere at most 3.5 degrees (= ) from a pointing center, a scale of 1 arcsecond per pixel, and the projection TAN.
Recently, the International Virtual Observatory Alliance (IVOA) agreed on a flexible and extensible framework for registering astronomical resources in a distributed virtual registry [IVOA 2003]. We expect to enhance the Hyperatlas standard by providing an extension schema for this registry. This would allow Hyperatlas-compliant imagery to be published and discovered through the Virtual Observatory, bringing to full flower the promise of image federation technology.
IVOA, International Virtual Observatory Registry Framework, 2003,
http://www.ivoa.net/twiki/bin/view/IVOA/IvoaResReg
Szalay, A. S., Connolly, A. J., Szokoly, G. P. 1998, AJ, 117, 68
Williams, R. D., The NVO Hyperatlas Standard, 2003a,
http://bill.cacr.caltech.edu/usvo-pubs/files/hyperatlas.pdf
Williams, R. D., Feldmann, M. T., Atlasmaker, Software to build Atlases of Federated Imagery, 2003,
http://www.cacr.caltech.edu/projects/nvo/atlasmaker/