Needles in a Haystack: Harnessing Machine Learning and Citizen Science to Catch Small Body Activity in Action

Active asteroids—hybrid objects that exhibit characteristics of both asteroids and comets—provide unique insights into solar system evolution and the current distribution of volatiles. However, their apparent rarity (~60 known), coupled with the petabyte-sized haystacks of archival survey data in which they may be hidden, makes detections challenging. This work explores the application of machine learning to constrain populations of active small bodies by increasing the rate of data evaluation. Specifically, we employ convolutional neural networks (CNNs) for object detection and classification, tailored to mirror the workflow of Citizen Scientist volunteers who classify objects of interest from thumbnails of archival data.

Trained on a dataset of labeled active small bodies and a large control set of inactive objects, our CNN evaluates image cutouts centered on known small bodies and classifies them based on signs of activity, such as tails or comae. We explore training datasets from both the Canada-France-Hawaii Telescope (CFHT) MegaCam and Subaru Hyper Suprime-Cam (HSC) archives, highlighting the adaptability of the CNN to be retrained and applied to different survey datasets. Our work also evaluates the influence of activity rate among training data in comparison to the predicted activity rate among Main Belt asteroids (1:10,000) and includes a robust response to erroneous thumbnail images.

We find that the CNN demonstrates high precision and recall across both CFHT and HSC archives and present our first results of the CNN’s high-confidence detections, whether previously known or not. This work illustrates the potential of machine learning techniques to accelerate discoveries of active small bodies and is intentionally designed to be used alongside proven Citizen Science applications. Combining AI with by-eye evaluation gives us a powerful and versatile tool in the doorway to next-generation surveys like LSST.