Sunday 6 March 2022

Using drones and machine learning to search for an observed meteorite in Australia's Western Nullarbor.

Meteorites can tell us much about the history of the Solar System, having formed in the protoplanetary nebula from which that system formed. They provide useful insights into the physical and chemical compositions of asteroids, and even larger Solar System bodies, but can be hard to find, and begin to be affected by processes on Earth from the moment they arrive. In recent years, fireball observatory networks have been set up in a number of areas, each of which uses a series of cameras to track meteor falls, enabling scientists to track their trajectories and therefore an approximate idea of where any meteorites might have fallen, as well as to calculate their original orbital paths, providing valuable insights into the connection between the composition of meteorites and their origin within the Solar System. However, locating meteorites is still a labourious process, usually involving a team of trained searchers walking in a line 5-10 m apart, sweeping the area of the fall until the meteorite is found, or, more often, the search is abandoned; only about 20% of such searches are successful.

In a paper published on the arXiv database at Cornell University on 3 March 2022, Seamus Anderson, Martin Towner, John Fairweather, Philip Bland, Hadrien Devillepoix, Eleanor Sansom, Martin Cupak, Patrick Shober, and Gretchen Benedix of the Space Science and Technology Centre at Curtin University present the results of a study which used drones and machine learning to search for an observed meteorite in Australia's Western Nullarbor Desert.

The observed meteor fell over the Lintos Paddock area of Kybo Station in Western Australia on the evening of 1 April 2021. This was observed by two cameras operated by Curtain University's Desert Fireball Network, located at Mundrabilla Station and O'Malley Siding, 149 km and 471 km to the east of the end point of the meteor's trajectory. The object was traced from an initial altitude of 87 km until it was only 25 km above the ground, during which time it slowed from 25.4 km per second to 8.4 km per second over a period of 3.1 seconds, while travelling on a slope of 64°. Because of the distance from the observation sites, and the lack of triangulation due to both sites being in the same direction, Anderson et al. created a series of models to predict the final end point of the object.

The DFN 09 meteorite fall at Kybo Station, Western Australia. (Clockwise from top) Fireball observations from DFN camera stations at Mundrabilla Station and O’Malley Siding, and their location within Western Australia; The 90% certainty searching area (transparent white), the best fit fall line (red markers), and the location of the recovered meteorite (yellow star); Pre-impact orbit for the DFN 09 meteoroid. Anderson et al. (2022).

Anderson et al. calculated that the object would have a mass of between 150 g and 700g, and identified an area 5.1 km², within which there was a 90% certainty of the meteorite having fallen. This was a high enough certainty to warrant a visit to the area, resulting in a three-day field trip, during which the area was surveyed with a drone and the data from the survey processed by machine learning, resulting in the eventual recovery of a 70 g meteorite.

Favourable and unfavourable prediction distributions from two images. Given a 70% confidence threshold, Image/Distribution (A) will return 3 meteorite candidates, while Distribution (B) will return over 100 candidates. (B)-like images are later used for retraining. For clarity, the number of detections displayed in image (B) is capped at 50. Anderson et al. (2022). 

A DJI M300 drone with a Zenmuse P1 camera was used to study the target area, producing 57 255 images with a 20% overlap. Of these, Anderson et al. were able to process 5096 on site, producing a total of 46 501 000 tiles for their machine learning algorithm to analyse, by comparing patterns to a database of known Nullarbor objects. From these tiles, the algorithm identified 56 384 first stage candidates, which were then studied with a 3x3 grid graphical user interface to eliminate obvious false positives. This reduced the number of candidate tiles to 259 second stage candidates, which were inspected with a second user interface, this time allowing toggling and zooming, to identify more likely objects. This reduced the number of candidates to 38, which were then revisited by the drone for closer inspection, finally reducing the number of candidate objects to four, which were then visited directly by Human researchers.

The four stage process for eliminating false positives and verifying meteorite candidates. (From Left to Right) (1) Grid GUI. (2) Zoom-pan GUI. (3) Drone visit. (4) In-person visit.

The meteorite was found less than 50 m from the ideal line predicted by Anderson et al. in the 88th image taken on by the third flight of the drone on the first day of the study. While it has not formally been studied or classified yet, it is a 70 g object measuring approximately 5 x 4 x 3 cm, with a preferentially smoothed side, and a fusion crust typical of chondritic meteorites that have passed through the atmosphere.

The recovered meteorite as seen in person (top two), and from the survey drone (bottom one). For scale, a 15 cm long felt pen is placed next to the meteorite (top right). The yellow box in the bottom image is 22 cm on one side. Anderson et al. (2022).

Although their software enabled them to find and recover the meteorite rapidly, Anderson et al. are at pains to emphasise that they have not developed a meteorite-detection program, but rather a program for detecting anomalous objects in images of an area, in this case the Western Nullarbor. As well as the meteorite, the program detected a range of other anomalies, including tin cans, bottles, Snakes, Kangaroos, and piles of bones from multiple Animals. They also note that the program detected equipment used by the surveyors and left around their camp, and which also had not been included in the training data given to the program. They hope that in future the software can be used not just to identify meteorites, but also in fields such as wildlife monitoring, or search and rescue.

See also...

Follow Sciency Thoughts on Facebook.

Follow Sciency Thoughts on Twitter.