Improving supernova detection by using YOLOv8 for astronomical image analysis

In the realm of astronomical imagery, the identification of supernovae poses a complex and intricate challenge. This intricacy extends beyond mere luminosity assessment, encompassing the discernment of diverse patterns inherent to the celestial phenomenon. Recent advancements in the field of computer vision have sought to address this challenge through the development of novel models. The labeled telescopic images capturing supernovae instances are collected from two distinct observatories, namely Pan-STARRS (Panoramic Survey Telescope and Rapid Response System) and PSP (Popular Supernova Project), strategically positioned at disparate global locations. In this paper, we delve into the application of the cutting-edge YOLOv8 (You Only Look Once) model for supernova detection. Specifically, in this study, a comparison was made with other state-of-the-art (SoTA) models over 80:20, 50:50, and 20:80 train-test ratios. YOLOv8 has a superior performance by obtaining 98.9%, 98.5%, and 96.9% \(\textrm{mAP}^{.50:.95}\) ${\text{mAP}}^{.50:.95}$ scores respectively. The presented values reveal the efficacy of YOLOv8 when applied to datasets featuring small-size bounding boxes, in the context of supernova detection. Hence, a noteworthy enhancement has been realized within the domain of astronomical imagery.