The combined power of citizen science and machine learning have led to the discovery of more than 1,000 new asteroids in archival images taken by the Hubble Space Telescope.
More than 1,000 previously unknown asteroids have been found on old Hubble photos — not by professional astronomers, but by citizen scientists and a machine-learning module. “This work is a great example of how serendipitous science can take place in large datasets,” comments asteroid researcher Bill Bottke (Southwest Research Institute), who was not involved in the study.
Over the past three decades, hundreds of thousands of deep-sky images have ended up in the Hubble archives. Occasionally, tiny asteroids passed through the field of view during the exposure, leaving a faint trail on the photo. Such asteroid trails are usually curved, because Hubble itself is also moving in its orbit.
In 1998 Robin Evans and Karl Stapelfeldt (both at Jet Propulsion Laboratory) unearthed dozens of “photobombing” asteroids by visually inspecting Hubble images. But since then, no one has carried out another dedicated search. “Evans and I have been too busy with other projects to keep up with the flow of new Hubble images to search,” says Stapelfeldt.
Sandor Kruk (Max Planck Institute of Extraterrestrial Physics, Germany) and Pablo García Martín (Autonomous University of Madrid, Spain) took a different approach. Building on Kruk’s experience with GalaxyZoo, a citizen-science program originally developed to help astronomers classify galaxies, they developed a similar program called Hubble Asteroid Hunter, and invoked the help of thousands of volunteers from all over the world.The number of known asteroids in the solar system is steadily increasing. This animation maps all known asteroids discovered between 1999 and 2018. The newly discovered asteroids in Hubble images add to this ongoing census.
Mining for Asteroids
Identifying asteroids on Hubble images is difficult. Most of the trails are exceedingly faint. Moreover, viewers must distinguish them from other elongated artifacts, such as satellite trails, cosmic ray tracks (caused by charged particles slamming into the detector), and faint gravitational lens arcs.
Kruk and his colleagues selected 37,323 Hubble images obtained over the past 20 years with Hubble’s Advanced Camera for Surveys and Wide Field Camera 3. They cut the images into four quadrants, each of which was inspected by 10 volunteers. Their task was to identify every possible track, mark the start and end points, and if possible, provide a classification.
By the end of the program, 11,482 volunteers had searched through 144,559 quadrants. Between them, they ended up with a whopping 1.78 million individual classifications (155 per person on average). By critically comparing and combining all these data, the team ended up with 1,488 potential asteroid trails – on average just one for every 100 image quadrants.
Meanwhile, Kruk’s team fed the reported classifications into a machine-learning module, developed in collaboration with Google scientists. After sufficient training, the algorithm rediscovered about two-thirds of the trails that the citizen scientists had found. In addition, it found almost 1,000 new trails in images that no one had looked at before.
“The availability and enthusiasm of citizen scientists were pivotal for the development of the automated algorithm,” says Stapelfeldt, who is “happy to see this work go forward.” Kruk agrees. “In principle, we could now let the algorithm search for asteroid trails in each and every new Hubble image,” he says. “This would never have been possible without the efforts of citizen scientists.”
But even with 2,487 trails on hand, the team wasn’t done yet. As they describe in a paper to appear in Astronomy & Astrophysics (preprint available here), Kruk and two other team members visually inspected each of them, weeding out remaining cosmic ray events that had not properly been recognized as well as false or ambiguous identifications. In the end, a total of 1,701 valid asteroid trails remained.
Next, the team checked the results against the existing database of small solar system objects hosted by the Minor Planet Center of the International Astronomical Union. In 670 cases, they identified the trail with a known object. For example, the relatively bright asteroid 2001 SE101 passed in front of the Crab Nebula on a Hubble image obtained in 2005. That left 1,031 unidentified trails, most of them very faint: 23rd-magnitude on average.
“To me, the most exciting thing is that this kind of science turns out to be possible at all with the Hubble data, despite the fact that Hubble was never designed to do this,” says Kruk.
Bottke is equally enthusiastic. “This paper shows there are many ways to maximize the science we can get from all telescopic observations,” he says, “and that entrepreneurial amateur astronomers can produce some really neat results.”
What to Do with 1,000 New Asteroids
In future work, Kruk and his colleagues will analyze the trail shapes of the new asteroids, which will enable them to derive distances. They can then translate the observed brightness of an object into an estimate of its physical size. Eventually, this will provide more precise information on the size distribution of the smallest objects in the solar system.
“We know the main [asteroid] belt is collisionally evolved, with collisions being the primary geologic process affecting asteroids,” says Bottke. “By probing the number of small bodies that exist, we get critical input for computer models that allow us to estimate both the strengths of asteroids and how long they are likely to survive prior to being disrupted. This tells us about asteroid evolution in general.”
René Laureijs (European Space Agency, ESA) looks forward to future cooperation with citizen scientists in ESA’s Euclid mission, for which he is the project scientist. Euclid will map two-thirds of the entire sky at the same resolution as Hubble to study dark matter and dark energy in the universe. “With Euclid, the main focus will be on the morphology of galaxies,” Laureijs says, “but we also expect to find at least 150,000 small solar system objects in our data.”