Data Representation Matters: Optimizing Machine Learning for X-ray Source Classification by Leveraging Spatio-Spectral Information

Abstract

Distinguishing between faint extended X-ray sources is a fundamental yet challenging task in X-ray surveys. This paper presents a benchmark comparison of convolutional neural network models for classifying simulated XMM-Newton observations of AGN and galaxy clusters, focusing on the impact of data representation and interpretability of that data. We evaluate 1D, 2D, 2D+1D, and 3D CNNs trained on distinct data representations derived from simulated XMM-Newton observations of AGN and clusters: 1D (spectral), 2D (imaging) and 3D (event cube) data. We further apply 3D-GradCAM to interpret the learned features and assess the physical basis for the model's decision making to move beyond black-box classification. Our results demonstrate that increased spectral resolution enhances classifier performance using 1D and 2D information provides a strong baseline, the integration of both domains yields significant diagnostic gains. We find that a 3D CNN architecture applied directly to spatio-spectral event cubes provides the most robust performance, with superior consistency across cross-validation folds. Interpretability analysis reveals that the 3D model autonomously learns to distinguish the diffuse thermal emission of the intracluster medium from the localized, non-thermal power-law signatures of AGN. It exploits multi-dimensional correlations that are lost in lower-dimensional projections. While the performance gain over combined 2D and 1D baselines is marginal, the 3D approach demonstrates an ability to perform implicit background characterisation. By operating directly on total photon counts within event cubes, this method bypasses the need for manual background subtraction and modelling, offering a consistent, streamlined, end-to-end pipeline for automated classification of X-ray sources in large-scale catalogues.