Galaxy Classification Using Artificial Neural Networks

Applications of machine-automated reduction procedures and pattern recognition techniques are a key ingredient to the execution of large all-sky digital surveys. The development of objective machine techniques for object detection, measurement, and classification not only allow for a timely reduction of the data, but greatly add to the scientific worth of the final product by imposing a reduction process which is rigorously adhered to. Works by Slezak et ul.,' Heydon-Dumbleton et ~ l . , ~ Rhee,3 and Odewahn et aL4 discuss the use of automated techniques for rudimentary image analysis (i.e., star-galaxy separation), but the area of automated morphological classification of galaxies is in a relative state of infancy. Early investigations have yet to produce a system which can discern the subtle features attainable by experienced human classifiers when assigning a type in the revised Hubble classification system. Using high-quality photometric parameter sets from multiple spectral regions, along with sophisticated classification algorithms, it is only a matter of time before machine classified sets will match the repeatability of the human classifiers. In addition, such techniques will be applied in a more consistent manner to vastly larger data sets than have heretofore been attempted in human visual classification projects. Thonnat' has used a rule-based expert system to derive galaxy classifications from parameters obtained in the image segmentation of PDS density arrays measured on photographic Schmidt plates. Although a well-understood classification system is attained in such a scheme, one must possess either a priori knowledge of how the measured image parameter data relate to the galaxy types or a vast amount of examples with which to derive the classification rules. Spiekermann6 has classified galaxies into five broad morphological classes (E, SO, Sa, Sb, Sc/Ir) using a fuzzy algebra system for digitized IIIa-J Schmidt images. Whitmore' presents a classification approach using a principal component analysis (PCA) of a variety of photometric parameters, some of which are based on metric measures, and hence require one to know the galaxy distance. A recent approach along this direction by Han,' where PCA is used to characterize the I-band luminosity profile of a galaxy, shows great promise. In recent work, Doi et aL9 have used the surface brightness and image concentration to discriminate between early and late type galaxies. Similarly, Abraham et al." use an image asymmetry and concentration parameter to perform a linear separation in a two-dimensional space for galaxy classification in three broad categories. From all these works, it is clear that one wishes to combine as