Type: Poster
Session: Posters (Wednesday & Thursday)
Author: Tatiana Acero-Cuellar

Abstract: The Vera C. Rubin Observatory will collect more than 500 images every night, each with 3.2 Gigapixels, approximately 20TB of data every night. Real astrophysical transients are rare events compared to the large number of artifacts or “bogus objects” generated by the survey and data preprocessing steps. Machine learning has long been used to address this problem (Goldstein et al. 2015) although a purely image-based approach is somewhat novel (Cabrera-Vives et al. 2016, Sedaghat et al. 2018, Acero-Cuellar et al. 2023). Deep learning (DL) techniques are well suited to address this task, these techniques vary in complexity, interpretability, and efficiency, with Convolutional Neural Networks being the most common. However, it is difficult to compare the performance reported in the literature once models are trained and implemented on different data sets from different telescopes and cameras and also using simulations generated following different approaches. This work provides an initial answer to which of the proposed DL-image-based models is best suited for the Real-Bogus classification of the upcoming Rubin data.