Study on cosmic ray background rejection with a 30 m stand-alone IACT using non-parametric multivariate methods in a sub-100 GeV energy range

Abstract

During the last decade ground-based very high-energy γ-ray astronomy achieved a remarkable advancement in the development of the observational technique for the registration and study of γ-ray emission above 100 GeV. It is widely believed that the next step in its future development will be the construction of telescopes of substantially larger size than the currently used 10 m class telescopes. This can drastically improve the sensitivity of the ground-based detectors for γ rays of energy from 10 to 100 GeV. Based on Monte Carlo simulations of the response of a single stand-alone 30 m imaging atmospheric Cherenkov telescope (IACT) the maximal rejection power against background cosmic ray showers for low energy γ-rays was investigated in great detail. An advanced Bayesian multivariate analysis has been applied to the simulated Cherenkov light images of the γ-ray-and proton-induced air showers. The results obtained here quantitatively testify that the separation between the signal and background images degrades substantially at low energies, and consequently the maximum overall quality factor can only be about 3.1 for γ rays in the 10-30 GeV energy range. Various selection criteria as well as optimal combinations of the standard image parameters utilized for effective image separation have been also evaluated.

Gamma/Hadron Separation in Imaging Air Cherenkov Telescopes Using Deep Learning Libraries TensorFlow and PyTorch

“…. This means that one of more elaborated versions of this technique [5,6,7,8,9] would definitely overwhelm our CNN result. Therefore, some image cleaning procedure is necessary before the CNN implementation; however, the one we used is not optimal for CNN quality.…”

“…A fully automatic algorithm without image cleaning demonstrated the same quality as the simplest version of the conventional technique, Hillas analysis [3]. This means that one of more elaborated versions of this technique [5,6,7,8,9] would definitely overwhelm our CNN result. Therefore, some image cleaning procedure is necessary before the CNN implementation; however, the one we used is not optimal for CNN quality.…”

“…For example, consecutive cuts on various Hillas parameters can be used to remove background events step-by-step, with optimal cut values determined by Monte Carlo simulation of the telescope [4]. Various attempts of combining these empirical parameters and derived quantities into machine learning algorithms have been carried out [5,6,7,8,9], but only three very recent efforts of a deep learning approach, which means fully automatic choice of the image features instead of empirical ones, have been made so far [10,11,12], of which the first one is an attempt not to distinguish primary gammarays from cosmic rays, but to select a special kind of images produced by secondary muons ('Muon Hunter' project [13]).…”

STEREO ARRAY of 30m imaging atmospheric Čerenkov telescopes: A next-generation detector for ground-based high energy gamma-ray astronomy

“…However none of them demonstrated drastically improved performance as compared with the standard multi-parametric approach ("super-cuts"). Recent studies on the multivariate analysis in application to the data simulated for a 30 m telescope have shown that a realistic quality-factor for such low energy events is of about Q ≃ 3, which corresponds to the γ-ray and cosmic ray acceptance of κ γ ≃ 0.3 and κ CR ≃ 0.01, respectively (Konopelko et al 2005). This was achieved using Bayesian decision rules and a non-parametric estimate of the multi-parameter density.…”