(P1.04) Bayesian Belief Networks for Astronomical Object Recognition and Classification in CTI-II

Mike Ritthaler ((University of New Mexico, Dept. of Computer Science)
George Luger (University of New Mexico, Dept. of Computer Science)
Robert Young (University of New Mexico, Dept. of Computer Science)
John McGraw (University of New Mexico, Physics & Astronomy Dept.)
Pete Zimmer (University of New Mexico, Physics & Astronomy Dept.)

The University of New Mexico (UNM) is currently designing and building the CCD Transit Instrument II (CTI-II), a 1.8m transit survey telescope. The stationary CTI-II uses the time delay and integrate readout mode for a mosaic of CCDs to generate over 200 gigapixels per night which is required to be analyzed within a day of observation. In order to assist the development of the needed processing system, a collaboration has been established between the Physics & Astronomy Department and the Computer Science department at UNM called the CTI-II Computing Collective (C3). C3 leverages a pool of Computer Science Master’s Degree students and their advisors to assist both in the development of the processing system and novel analysis techniques, all while training a cadre of graduates with knowledge and experience in image processing.
For processing the image data generated by survey telescopes, the automated location and classification of objects in the field becomes very important for both real-time analysis and near real-time processing. Because many instruments, such as CTI-II, operate within a wide range of both seeing and sky-brightness conditions, we show an application of Bayesian Belief Networks (Heckerman 1999, Friedman & Goldszmidt 1996) in astronomical object classification that takes into account those conditions. Bayesian Belief Networks analyze the relationships between both discrete and continuous random variables in a graphical representation where dependencies are represented by directed arcs in the network and random variables by nodes. By using nodes to represent seeing, sky brightness, and statistical invariants of objects found in the field, we present results using a Bayesian Belief network as a naive Bayesian classifier, a tree-augmented Bayesian Belief Network structure, and an unconstrained structure on both real and synthetic data. These results are benchmarked against Source Extractor (Bertin & Arnouts 1996) on the same data. By ordering the dependency arcs, we show how the network can be used for prioritizing analysis methods so that the most effective methods can be applied first. Finally, we suggest how a network can be used for meta-analysis of other techniques, including both supervised and unsupervised learning algorithms as well as other classification methods.