SKICAT: A Machine Learning System for Automated Cataloging of Large Scale Sky Surveys

Abstract

We describe the application of decision tree based classification techniques to the development of an automated tool for the reduction of a large scientific data set. The 2nd Palomar Observatory Sky Survey (POSS-II) will provide a comprehensive coverage of the northern celestial hemisphere in the form of digitized photographic plates containing on the order of 0.5 billion objects. It is necessary to classify the detected objects as stars, galaxies, or instrumental artifacts. The size of this data set precludes manual analysis. Image processing routines are used to identify sky objects and to extract a set of basic features for each object that are used to select a useful and robust set of attributes for classifying sky objects. The GID3* and O-BTree decision tree learning algorithms, in conjunction with the RULER system for statistically pruning and merging multiple trees, are used to classify the detected objects. The learning algorithms produced classifiers that can classify objects in the survey that are too faint for visual classification with an accuracy level of about 94%. This increases the number of classified objects in the final catalog by three-fold relative to the best techniques previously employed by extending at least one magnitude fainter than objects classified in photographic sky surveys to date. SKICAT represents a system in which machine learning played a powerful and enabling role, and solved a difficult, scientifically significant problem.