Correntropy Based Filtering for Supernova Detection

The recent surge of huge astronomical telescopes opens new possibilities for astrophysical studies, with corresponding challenges regarding massive data analysis. Taking advantage of these new technologies, the High Cadence Transient Survey (HiTS) has developed a pipeline to detect transient astronomical phenomena, such as supernovae, by analyzing pairs of stamps of objects extracted directly from the telescope images in an online fashion. We introduce a new complementary method for detecting faint, low signal-to-noise ratio supernovae by discriminating the evolution of luminosities of pixels through filtering sequences of images. Our method uses the Correntropy Filter (CF), which is a nonlinear extension of the classic Kalman Filter. CF uses an information-theoretic cost function called Cross-Correntropy that allows the discarding of artifacts and outliers, which are abundant in astronomical measurements. The proposed method was tested on the HiTS surveys and the results show that it is able to rediscover most supernovae in the high cadence epochs. In addition, it has found 30 new supernova candidates too faint to be discriminated in a pair of stamps. The analysis of longer sequences of images allows us to reduce the number of false positives, but at the expense of a late detection.