Using deep learning and crowdsourcing to survey asteroid trails in ESA's Hubble data archive

The Hubble Space Telescope (HST) archives hide many unexpected treasures, such as trails of asteroids, showing a characteristic curvature due to the parallax induced by the orbital motion of the spacecraft. We have explored two decades of HST data for serendipitously observed asteroid trails with a deep learning algorithm on Google Cloud and trained on classifications from the Hubble Asteroid Hunter (www.asteroidhunter.org) citizen science project.  The project was set up as a collaboration between the ESAC Science Data Centre, Zooniverse, and engineers at Google as a proof of concept to valorize the rich data in the ESA archives. I will present the first results from the project, finding 1700 asteroid trails in the HST archives (Kruk et al., 2022). The majority of the trails we found are faint (typically > 22 mag) and correspond to previously unidentified asteroids. Identifying the asteroids in HST images allows us to refine their ephemerides, as well as study their orbital distribution. I will argue that a combination of AI and crowdsourcing is an efficient way of exploring increasingly large datasets by taking full advantage of the intuition of the human brain and the processing power of machines. This comprehensive survey of asteroids required only ten hours of wall clock time to train the algorithm and scan the entire HST archival dataset, illustrating the immense scientific potential of these techniques and the benefits of exploiting the vast amounts of data in the archives.