The mounting data stream of large time-domain surveys renders the visual inspections of a huge set of transient candidates impractical. Techniques based on deep learning-based are popular solutions for minimizing human intervention in the time domain community. The classification of real and bogus transients is a fundamental component in real-time data processing systems and is critical to enabling rapid follow-up observations. Most existing methods (supervised learning) require sufficiently large training samples with corresponding labels, which involve costly human labeling and are challenging in the early stages of a time-domain survey. One method that can make use of training samples with access to only a limited amount of labels is highly desirable for future large time-domain surveys. These include the forthcoming 2.5-meter Wide-Field Survey Telescope (WFST) six-year survey and the Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST). Deep-learning-based methods have been favored in astrophysics owing to their adaptability and remarkable performance. They have been applied to the task of the classification of real and bogus transients. Unlike most existing approaches, which necessitate massive and expensive annotated data, we aim to leverage training samples with only 1000 labels and discover real sources that vary in brightness over time in the early stages of the WFST six-year survey. We present a novel deep learning method that combines active learning and semi-supervised learning to construct a competitive real-bogus classifier. Our method incorporates an active learning stage, where we actively select the most informative or uncertain samples for annotation. This stage aims to achieve higher model performance by leveraging fewer labeled samples, thus reducing annotation costs and improving the overall learning process efficiency. Furthermore, our approach involves a semi-supervised learning stage that exploits the unlabeled data to enhance the model's performance and achieve superior results, compared to using only the limited labeled data. Our proposed methodology capitalizes on the potential of active learning and semi-supervised learning. To demonstrate the efficacy of our approach, we constructed three newly compiled datasets from the Zwicky Transient Facility (ZTF), achieving average accuracies of 98.8<!PCT!>, 98.8<!PCT!>, and 98.6<!PCT!> across these three datasets. It is important to note that our newly compiled datasets only work in terms of testing our deep learning methodology and there may be a potential bias between our datasets and the complete data stream. Therefore, the observed performance on these datasets cannot be assumed to directly translate to the general alert stream for general transient detection in actual scenarios. The algorithm will be integrated into the WFST pipeline, enabling an efficient and effective classification of transients in the early period of a time-domain survey.
Read full abstract