Object classification in astronomical multi-color surveys

We present a photometric method for identifying stars, galaxies and quasars in multi-color surveys, which uses a library of greater than or similar to 65000 color templates for comparison with observed objects. The method aim s for extracting the information content of object colors in a statisticall y correct way, and performs a classification as wed as a redshift estimatio n for galaxies and quasars in a unified approach based on the same probabil ity density functions. For the redshift estimation, we employ an advanced v ersion of the Minimum Error Variance estimator which determines the redshif t error from the redshift dependent probability density function itself. Th e method was originally developed for the Calar Alto Deep Imaging Survey (C ADIS), but is now used in a wide variety of survey projects. We checked its performance by spectroscopy of CADIS objects, where the method provides hi gh reliability (6 errors among 151 objects with R < 24), especially for the quasar selection, and redshifts accurate within <sigma>(z) approximate to 0.03 for galaxies and sigma (z) approximate to 0.1 for quasars. For an opti mization of future survey efforts, a few model surveys are compared, which are designed to use the same total amount of telescope time but different s ets of broad-band and medium-band filters. Their performance is investigate d by Monte-Carlo simulations as well as by analytic evaluation in terms of classification and redshift estimation. If photon noise were the only error source, broad-band surveys and medium-band surveys should perform equally well, as long as they provide the same spectral coverage. In practice, medi um-band surveys show superior performance due to their higher tolerance for calibration errors and cosmic variance. Finally, we discuss the relevance of color calibration and derive important conclusions for the issues of lib rary design and choice of filters. The calibration accuracy poses strong co nstraints on an accurate classification, which are most critical for survey s with few, broad and deeply exposed filters, but less severe for surveys w ith many, narrow and less deep filters.