Supernovae Detection with Fully Convolutional One-Stage Framework.

Area of Science:

• Astronomy and Astrophysics
• Computer Science
• Data Science

Background:

• Astronomical sky surveys generate massive datasets, pushing astronomy into a big data era.
• Manual identification of supernovae from this data is challenging due to volume and rarity.
• Existing deep learning methods often use simulated data or lack generality for real-world applications.

Purpose of the Study:

• To evaluate state-of-the-art object detection algorithms on real supernova data.
• To develop an improved detection algorithm for supernovae.
• To address limitations in existing methods, such as assumptions about image cropping and single-candidate localization.

Main Methods:

• Collected and organized data from the Panoramic Survey Telescope and Rapid Response System (Pan-STARRS) and the Popular Supernova Project (PSP).
• Compared several object detection algorithms on these real datasets.
• Selected the Fully Convolutional One-Stage (FCOS) method as a baseline and enhanced it with data augmentation, attention mechanisms, and small object detection techniques.

Main Results:

• Demonstrated the performance of various detection algorithms on real supernova datasets.
• Achieved significant performance enhancements with the improved FCOS-based detection algorithm.
• Validated the effectiveness of data augmentation, attention mechanisms, and small object detection techniques for this task.

Conclusions:

• The developed FCOS-based algorithm shows superior performance for supernova detection using real astronomical data.
• The new datasets and improved methods advance the field of automated supernova discovery.
• This work provides a more robust approach for handling large-scale astronomical data in supernova research.