Automatic Detection of Exoplanet Transit Signals

Automatic Classification of Kepler Threshold Crossing Events

Authors:

McCauliff et al

Abstract:

The Kepler Science Operations Center detects interesting, exoplanet transit-like signals while searching over 211,000 distinct light curves. The mission has produced four catalogs of interesting objects with planet transit-like features known as Kepler Objects of Interest (KOI). The total number of objects with transit-like features identified in the light curves has increased to as many as approximately 18,000, just examining the first three years of data. This number of significant detections has become difficult for human beings to inspect by eye in a thorough and timely fashion. In order to accelerate the process by which new planet candidates are classified and to provide an independent assessment of planet candidates, we propose a machine learning approach to establish a preliminary list of planetary candidates ranked from most credible to least credible. The classifier must distinguish between three classes of detections: non-transiting phenomena, astrophysical false positives, and planet candidates. We use random forests, a supervised classification algorithm, that has an error rate of 1.34 percent with some qualifications.