Presentation #318.05 in the session Computation, Data Handling, Image Analysis II.
Detecting cosmic rays (CRs) in single-exposure images has been challenging, especially in data from different ground-based instruments with variable conditions. We present Cosmic-CoNN, a deep-learning CR detector deployed at the Las Cumbres Observatory (LCO) that achieves 99.91% true-positive detection rate on imaging data and 97.4% on spectroscopic data with a false-positive rate of 0.01%. More importantly, our model is well generalized to other ground-based instruments. It reaches over 96.40% true-positive rates on imaging data from Gemini Observatory’s GMOS-N/S telescopes while we never used Gemini data during training. In this talk, we will demonstrate: 1. Cosmic-CoNN’s interactive interface for CR detection, visualization, and pixel-level mask editing. 2. The command-line interface and Python APIs for custom data-workflow integration, and 3. Other detection tasks that can benefit from our deep-learning framework (e.g., satellite or Starlink trails detection) and the dataset of 4,500 consecutive scientific observations from over 20 telescopes around the world. Our codebase, dataset, and tools are open-source and available at https://github.com/cy-xu/cosmic-conn.