NEURAL-NETWORK CLASSIFICATION OF THE NEAR-INFRARED SPECTRA OF A-TYPE STARS

We present an atlas of near-infrared (NIR) spectra of A stars for lumi nosity classes Ia through V in the 15 Angstrom resolution system descr ibed by Torres-Dodgen and Weaver (1993) and demonstrate an accurate me thod to automatically classify A stars on this system. Using equivalen t widths, artificial neural networks (ANNs) can classify these spectra to an accuracy of 0.4 types (subclasses) in temperature and 0.15 clas ses in luminosity. Using the spectrum, with no manual intervention exc ept wavelength registration, ANNs can classify these spectra with an a ccuracy comparable to that of 2 Angstrom resolution MK classification: 0.5 types in temperature and 0.35 classes in luminosity. In addition, ANNs can concurrently determine reddening to an accuracy of 0.05 in E (B-V). We demonstrate that this NIR-ANN spectral classification system has the primary properties needed for automated classification survey s: it is based in the most efficient spectral region of modern silicon -based detectors, it requires low-resolution (15 Angstrom) spectra to achieve sub-classification box accuracy, it can produce two-dimensiona l classifications at least as accurate as those by expert human classi fiers, it is relatively insensitive to interstellar reddening and can accurately determine the reddening, it can identify and classify compo site spectra, it degrades slowly with decreasing signal-to-noise ratio , and it requires a minimum of human interaction at all stages of the process.