Deep Representation Learning for Transient Event Detection in Astronomical Imaging

Abstract

Modern astronomical surveys generate increasingly large volumes of imaging data, creating significant challenges for the timely identification of transient celestial events. Traditional detection pipelines often struggle to distinguish astrophysical signals from imaging artifacts, atmospheric interference, and instrumental noise. This study proposes a deep representation learning framework for automated transient event detection using multi-epoch astronomical images. The proposed architecture combines convolutional feature extraction with cross-epoch attention to identify localized intensity variations while accounting for observational uncertainty. The illustrative evaluation employs 6.3 million astronomical image cutouts, including 41,000 labeled transient-event candidates obtained from a mixture of archival observations and simulated data. The proposed model achieves an AUROC of 0.976, outperforming a conventional convolutional classifier with an AUROC of 0.951. At a fixed sensitivity of 90%, the false-positive rate decreases from 2.1% to 0.8%, substantially reducing the number of non-astrophysical candidates requiring manual inspection. Average inference time is approximately 14 milliseconds per image cutout. These findings suggest that deep representation learning can improve the efficiency of automated transient identification in large-scale astronomical surveys.