Leveraging Movement Representation from Contrastive Learning for Asteroid Detection

Abstract

To support asteroid-related studies, current motion detectors are utilized to select moving object candidates based on their visualizations and movements in sequences of sky exposures. However, the existing detectors encounter the manual parameter settings which require experts to assign proper parameters. Moreover, although the deep learning approach could automate the detection process, these approaches still require synthetic images and hand-engineered features to improve their performance. In this work, we propose an end-to-end deep learning model consisting of two branches. The first branch is trained with contrastive learning to extract a contrastive feature from sequences of sky exposures. This learning method encourages the model to capture a lower-dimensional representation, ensuring that sequences with moving sources (i.e., potential asteroids) are distinct from those without moving sources. The second branch is designed to learn additional features from the sky exposure sequences, which are then concatenated into the movement features before being processed by subsequent layers for the detection of asteroid candidates. We evaluate our model on sufficiently long-duration sequences and perform a comparative study with detection software. Additionally, we demonstrate the use of our model to suggest potential asteroids using photometry filtering. The proposed model outperforms the baseline model for asteroid streak detection by +7.70% of f1-score. Moreover, our study shows promising performance for long-duration sequences and improvement after adding the contrastive feature. Additionally, we demonstrate the uses of our model with the filtering to detect potential asteroids in wide-field detection using the long-duration sequences. Our model could complement the software as it suggests additional asteroids to its detection result.