Cosmic-CoNN: A Cosmic Ray Detection Deep-Learning Framework, Dataset, and Toolkit

Abstract

Rejecting cosmic rays (CRs) is essential for scientific interpretation of CCD-captured data, but detecting CRs in single-exposure images has remained challenging. Conventional CR-detection algorithms require tuning multiple parameters experimentally, making it hard to automate across different instruments or observation requests. Recent work using deep learning to train CR-detection models has demonstrated promising results. However, instrument-specific models suffer from performance loss on images from ground-based facilities not included in the training data. In this work, we present Cosmic-CoNN, a deep-learning framework designed to produce generic CR-detection models. We build a large, diverse ground-based CR dataset leveraging thousands of images from the Las Cumbres Observatory global telescope network to produce a generic CR-detection model which achieves a 99.91% true-positive detection rate and maintains over 96.40% true-positive rates on unseen data from Gemini GMOS-N/S, with a false-positive rate of 0.01%. Apart from the open-source framework and dataset, we also build a suite of tools including console commands, a web-based application, and Python APIs to make automatic, robust CR detection widely accessible by the community of astronomers.

Cosmic Ray Rejection with Attention Augmented Deep Learning

“…Recent works, including deepCR (Zhang and Bloom, 2020), MaxiMask (Paillassa et al, 2020) and Cosmic-CoNN (Xu et al, 2021a) presented the CR detection as a supervised image segmentation problem and demonstrated the superiority of deep learning models on segmentation tasks. However, training deep learning algorithms in a supervised setting requires large amounts of annotated data.…”

“…We also considered the Las Cumbres Observatory (LCO) CR data set (Xu et al, 2021b), that has been made publicly available from Xu et al (2021a) to validate our CR detection models on previously unseen data. This data set has been constructed by leveraging the data from LCO's BANZAI data pipeline (Mc-Cully et al, 2018).…”

“…All the images are with an exposure time of 100 seconds or longer. For our analysis, we have used the test data from LCO CR dataset from three distinct imagers, with specifications specified in Table 1 of Xu et al (2021a). The CCDs used in these imagers are described in Brown et al (2013) and have pixel sizes in the range 9-15 µ m. The median seeing for these imagers is equal to 2.07", 2.04", and 1.44" for SBIG 6303, Sinistro, and Spectral imagers, respectively.…”

“…However, this approach is not suited for designing a generic CR detection model that can be used with a wide range of ground or space-based instruments and data statistics. Cosmic-CoNN (Xu et al, 2021a) presented a symbiotic combination of three enhancements to address the difficulties mentioned above. Based on the effective U-shaped architecture, Cosmic-CoNN proposed a novel loss function that explicitly handles the problem of class disparity, improved data sampling methods that have proven to be critical for training efficiency, and a feature normalization technique that uses group normalisation (Wu and He, 2018) to improve model generalization ability.…”

“…We note however that this ratio could also increase with the exposure time. A novel CNN-based algorithm called Cosmic-CoNN (Xu et al, 2021a) was proposed to develop generic CR detection models for all ground and spacebased instruments separately, while explicitly addressing the class imbalance problem. They also developed novel strategies for optimizing the neural network for astronomical images' unique spatial and numerical properties.…”