Application of pattern spectra and convolutional neural networks to the analysis of simulated Cherenkov Telescope Array data

TL;DR

This paper explores using pattern spectra derived from simulated Cherenkov Telescope Array data to train convolutional neural networks for gamma-ray energy reconstruction, aiming for faster, less resource-intensive analysis with minimal performance loss.

Contribution

The study introduces a novel approach of training CNNs on pattern spectra instead of raw images for CTA data analysis, improving efficiency while maintaining accuracy.

Findings

Pattern spectra enable effective energy reconstruction.

Training CNNs on pattern spectra reduces computational load.

Performance is comparable to traditional methods.

Abstract

The Cherenkov Telescope Array (CTA) will be the next generation gamma-ray observatory and will be the major global instrument for very-high-energy astronomy over the next decade, offering 5 - 10 x better flux sensitivity than current generation gamma-ray telescopes. Each telescope will provide a snapshot of gamma-ray induced particle showers by capturing the induced Cherenkov emission at ground level. The simulation of such events provides images that can be used as training data for convolutional neural networks (CNNs) to determine the energy of the initial gamma rays. Compared to other state-of-the-art algorithms, analyses based on CNNs promise to further enhance the performance to be achieved by CTA. Pattern spectra are commonly used tools for image classification and provide the distributions of the shapes and sizes of various objects comprising an image. The use of relatively shallow CNNs on pattern spectra would automatically select relevant combinations of features within an image, taking advantage of the 2D nature of pattern spectra. In this work, we generate pattern spectra from simulated gamma-ray events instead of using the raw images themselves in order to train our CNN for energy reconstruction. This is different from other relevant learning and feature selection methods that have been tried in the past. Thereby, we aim to obtain a significantly faster and less computationally intensive algorithm, with minimal loss of performance.