Convolution and Graph-based Deep Learning Approaches for Gamma/Hadron Separation in Imaging Atmospheric Cherenkov Telescopes

Abhay Mehta, Daniel R. Parsons, T. L. Holch, D. Berge, Matthias Weidlich2025 · DOI

TL;DR

This paper introduces novel deep learning architectures, specifically hybrid Convolutional and Graph Neural Networks (CNN-GNNs), for distinguishing gamma-ray events from background hadron showers in ground-based gamma-ray astronomy. By incorporating inductive biases that respect the physical structure of telescope data, these models show improved performance and offer a path towards more robust event classification in real-world observations.

What Problem Does This Paper Address?

Ground-based gamma-ray astronomy relies on Imaging Atmospheric Cherenkov Telescopes (IACTs) to detect faint gamma-ray signals. A major challenge is separating these signals from the overwhelming background of hadron-induced air showers, which can outnumber gamma-ray events by a factor of $10^4$. Traditional methods often use parameterized image features, but these struggle to fully exploit the complex correlations within the data. Deep learning offers a promising avenue by directly learning from image data, but existing models face challenges with robustness and generalization to real-world observational conditions, which differ significantly from simulations. This paper aims to address this by developing deep learning models with architectures that better align with the physical structure of IACT data.

Methods

The authors propose and evaluate three deep learning models for gamma/hadron separation:

  • Fast CNN-GNN: A hybrid model combining a Convolutional Neural Network (CNN) to extract features from individual telescope images with a Graph Neural Network (GNN) to aggregate information across telescopes. The GNN also incorporates telescope positions and basic image intensity features.
  • Strong CNN-GNN: An enhanced version of the Fast CNN-GNN that further enriches the graph representation by including reconstructed information about the air shower's core position. This is achieved by shifting the coordinate frame to center on the reconstructed impact point.
  • Hillas GNN: A baseline GNN model trained solely on traditional Hillas parameters extracted from the camera images, serving as a comparison to existing methods.

The CNN component uses an Inception-inspired architecture for multi-scale feature extraction, followed by fully connected layers. The GNN component employs an EdgeConv architecture with residual connections. Both CNN and GNN components are trained end-to-end.

Datasets

The models were trained and evaluated on simulated data from the High Energy Stereoscopic System (H.E.S.S.) array. This dataset consists of diffuse proton (hadron) and gamma-ray events simulated at a zenith angle of 20 degrees. The dataset was split into training (150,000 events, evenly split between gamma-ray and proton), validation, and testing sets. A secondary graph dataset using only reconstructed parameterized information (Hillas parameters) was also created for the baseline model.

Key Insights and Ideas

  • Inductive Biases Matter: The paper strongly advocates for designing deep learning architectures with inductive biases that mirror the physical structure of the data. For IACTs, this means leveraging the fact that multiple telescopes observe the same event, and the relationships between them are spatial and not necessarily ordered.
  • GNNs for Inter-Telescope Correlations: Graph Neural Networks are well-suited for aggregating information from multiple telescopes, as they can naturally handle varying numbers of nodes (telescopes) and incorporate specific node features.
  • Reconstructed Information Enhances Performance: Including physically meaningful reconstructed information, such as the shower core position, into the graph representation for the GNN leads to improved performance, particularly in rejecting background events.
  • Hybrid CNN-GNN Synergy: Combining CNNs for image-level feature extraction with GNNs for inter-telescope context provides a powerful approach that outperforms traditional methods and baseline GNNs.

Key Results

  • All proposed models successfully learned to classify gamma-ray and proton events.
  • The CNN-GNN models (Fast and Strong) outperformed the baseline Hillas GNN, achieving higher Area Under Curve (AUC) values on Receiver Operating Characteristic (ROC) curves.
  • The Strong CNN-GNN, which incorporates reconstructed core position, showed a greater ability to reject proton events compared to the Fast CNN-GNN. At a decision threshold of 0.5, this resulted in an additional 9% reduction in the False Positive Rate.
  • Permutation feature importance analysis revealed that features like "CNN-score," "Telescope position," and "Max. pixel intensity" were important for the models' decisions, with the specific importance varying between models and training instances.

Limitations and Future Work

  • Simulation vs. Reality Gap: The authors acknowledge that the primary limitation is the performance on real observational data, which is significantly more diverse and subject to physical uncertainties than simulated data. The current evaluation is based solely on simulations.
  • Generalization to Real Data: While the models show promise, their true test will be their performance on actual IACT data.
  • Future Directions: The authors suggest evaluating these models on real observations to definitively assess their viability and robustness. They also propose expanding training datasets to include events from a wider range of pointings and incorporating additional contextual information (e.g., zenith, azimuth) to further strengthen classification and generalization capabilities. The potential for extending the framework to heterogeneous telescope arrays is also highlighted.

Summary updated

Abstract

The identification of $\gamma$-rays from the predominant hadronic-background is a key aspect in their ground-based detection using Imaging Atmospheric Cherenkov Telescopes (IACTs). While current methods are limited in their ability to exploit correlations in complex data, deep learning-based models offer a promising alternative by directly leveraging image-level information. However, several challenges involving the robustness and applicability of such models remain. Designing model architectures with inductive biases relevant for the task can help mitigate the problem. Three such deep learning-based models are proposed, trained, and evaluated on simulated data: (1) a hybrid convolutional and graph neural network model (CNN-GNN) using both image and graph data; (2) an enhanced CNN-GNN variant that incorporates additional reconstructed information within the graph construction; and (3) a graph neural network (GNN) model using image moments serving as a baseline. The new combined convolution and graph-based approach demonstrates improved performance over traditional methods, and the inclusion of reconstructed information offers further potential in generalization capabilities on real observational data.
BibTeX
@misc{mehta2025convolution,
  author = {Abhay Mehta and Daniel R. Parsons and T. L. Holch and D. Berge and Matthias Weidlich},
  title = {{Convolution and Graph-based Deep Learning Approaches for Gamma/Hadron Separation in Imaging Atmospheric Cherenkov Telescopes}},
  year = {2025},
  doi = {10.22323/1.501.0752},
}
Figure