Signal-background separation and energy reconstruction of gamma rays using pattern spectra and convolutional neural networks for the Small-Sized Telescopes of the Cherenkov Telescope Array

TL;DR

This study explores using pattern spectra as input to CNNs for gamma-ray signal-background separation and energy reconstruction in CTA SSTs, finding it reduces computational costs but with some performance trade-offs.

Contribution

It introduces a pattern spectra-based approach for CNN input in gamma-ray analysis, offering a computationally efficient alternative to direct image training.

Findings

Pattern spectra reduce training computational cost by a factor of three.

Pattern spectra-based CNNs underperform compared to CNNs trained on raw CTA images.

Additional information beyond pattern spectra is needed for optimal CTA image analysis.

Abstract

Imaging Atmospheric Cherenkov Telescopes (IACTs) detect very-high-energy gamma rays from ground level by capturing the Cherenkov light of the induced particle showers. Convolutional neural networks (CNNs) can be trained on IACT camera images of such events to differentiate the signal from the background and to reconstruct the energy of the initial gamma ray. Pattern spectra provide a 2-dimensional histogram of the sizes and shapes of features comprising an image and they can be used as an input for a CNN to significantly reduce the computational power required to train it. In this work, we generate pattern spectra from simulated gamma-ray and proton images to train a CNN for signal-background separation and energy reconstruction for the Small-Sized Telescopes (SSTs) of the Cherenkov Telescope Array (CTA). A comparison of our results with a CNN directly trained on CTA images shows that the pattern spectra-based analysis is about a factor of three less computationally expensive but not able to compete with the performance of an CTA image-based analysis. Thus, we conclude that the CTA images must be comprised of additional information not represented by the pattern spectra.