ランダムフォレスト機械学習を用いた Tomo-e Gozen突発現象アラートシステムの開発
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
At the Kiso Observatory of the University of Tokyo, we conduct a high-cadence, wide-field optical survey every night using Tomo-e Gozen, a wide-field mosaic CMOS camera mounted on the 1.05-meter Schmidt telescope, detecting approximately 10 transient candidates per night. However, a major challenge arises from the contamination of over 100,000 false detections per night, primarily due to failures in the image subtraction process. To address this, Tomo-e Gozen has reduced the number of false detections by employing an image-based CNN classification model. We focused on methods that utilize astronomical catalogs and observational data with the aim of improving the accuracy for realizing real-time transient alerts. In this study, we developed a feature-based random forest model connected in series to the CNN model. As a result, we successfully constructed a Real/Bogus discriminator achieving a true positive rate of about 80 % and a false positive rate of about 5 %. Consequently, we reduced the number of false detections to only a few events per night. Additionally, we developed a random forest-based classifier for transient categorization, enabling the output of classification probabilities into types such as supernovae, cataclysmic variables, and flare stars for each detected transient candidate.
Fields of Study & Subjects
Gamma-ray bursts and supernovae
Astronomical Observations and Instrumentation
Astronomy and Astrophysical Research
Publication details
- Journal / source
- Institutional Repositories DataBase (IRDB)
- Volume
- JAXA-RR-25-003
- Pages
- 37–48
- Publication date
- 17 February 2026
- Language
- JA