Understanding planetary bodies will allow scientists to better understand Earth. With this knowledge, measures to preserve our planet can be taken. In 2015, NASA, the SETI Institute, and several other international space science organizations established the NASA Frontier Development Lab (FDL). Based out of the SETI Institute’s headquarters in Mountain View, CA, FDL uses AI (Artificial Intelligence) technologies to fill in gaps of understanding space study.

Postdoctoral student Robert Citron from the University of California Davis joined FDL to look into how meteorites that have made their way to Earth can fill in some of these gaps. What asteroid did the meteorite come from, how was the celestial body formed, what impacts its trajectory, and so much more can be gathered by studying meteorites. “In order to better understand the composition of asteroid families,” Robert states, “we much [sic] connect fresh meteorite falls to their pre-impact orbits. This is done by imaging the trajectories of the meteorites as they enter the atmosphere and finding the corresponding fragments.” But finding these meteorites, regardless of their size, can be a difficult undertaking.

Scientists can track meteorites coming into Earth’s atmosphere and determine the general area where they will land. But once these space rocks are on the ground, finding them can take hours, if not days to do. It requires manually searching wide swaths of land in which the meteorites could be partially buried in the ground, under shrubbery, or camouflaged by other rocks. “As part of the 2016 NASA Frontier Development Lab,” Robert went on to say, “we studied if machine learning could be used to help identify meteorite fragments in the field. We developed a framework using machine learning to classify images from a quadcopter drone to determine if the images contained meteorite fragments.”

As emergency workers around the world have learned, drones have become vital tools in locating missing people. The same principle being used to locate a person can be applied to locating hidden objects, like meteorite fragments on the ground. As drones are equipped with special sensors and high resolution cameras, they can gather visual data to be interpreted without people needing to manually walk through a field to search for elusive meteorite pieces.

To test if drones could be useful in finding meteorite fragments, Robert and his colleagues from UC Davis and FDL headed out to a suspected meteorite debris field near Walker Lake, Nevada. Before launching the drone, the team programmed it to fly at a set height ratio to the ground along a specific grid pattern. The drone then simply autonomously carried out the programmed mission. As the drone flew along the determined course, it collected a continuous stream of top-down images. To have manually searched the grid area, Robert explained, could have taken up to 100 hours. If the team was lucky, they may have found a single fragment in that time.

Instead, Robert and his team used an AI computer program to analyze the images that the drone rapidly collected. Robert went on to state that “These images can be spliced and fed into a machine learning object detection classifier, which can determine the likelihood that meteorite fragments are present in the image.” Robert and his team had to create a program that could self-recognize meteorites as foreign objects in the debris field. In total, Robert and his team conducted 10 test flights in the Walker Lake region. The drone, equipped with a GoPro camera took between 129-388 full images during each test flight.

Each of the images were analyzed by the AI program, called RetinaNet. If the AI program found a possible meteorite in an individual image, that image was marked as “positive” for further investigation. Any images without possible fragments, the program marked as “negative”. In the tests, the team placed decoy meteorites in the field to see just how well the drone and AI program could work in tandem. In the paper Robert released on the experiment’s findings, he said that there are still some adjustments to the AI system that need to be made for a fully viable program. Still, using the drone was a success primarily because it dramatically reduces the amount of time spent in the field. As part of the paper’s closing statement Robert said, “We demonstrated that as a proof of concept, it is possible to identify meteorites in the field by applying an object detection classifier to images acquired with an autonomous drone.”