86th Annual Meeting of the Meteoritical Society (2023) 6187.
pdf
AN UPDATE ON SEARCHING FOR METEORITES USING DRONES AND MACHINE LEARNING.
S. L. Anderson1, M. C. Towner1, E. K. Sansom1,2, H. A. R. Devillepoix1,2, M. Cupak1, 1Space Science and
Technology Centre, School of Earth and Planetary Science, Curtin University, GPO Box U1987, Perth, WA,
Australia (seamus.l.anderson@gmail.com), 2International Centre for Radio Astronomy Research, Curtin University,
GPO Box U1987, Perth, WA, Australia.
Introduction: As fireball networks become more common and widespread in their coverage [1], the need for
expedited meteorite recovery continues to grow [2, 3, 4, 5]. We report on the continuing efforts to recover
meteorites using drones and machine learning, specifically how this proven methodology [3] can be applied to
strewn fields. We also discuss our new strategy for combating false positives while retaining a high confidence for
meteorite detection.
Methods: Continuing from previous works [2, 3], we use the modules keras and tensorflow, accessed in python,
to train a convolutional neural network to identify subsections of images from a drone survey where meteorites may
be located. In this updated approach we automatically identify false positives by setting a confidence threshold, and
a detection threshold per image. If more than 10 ‘positives’ appear in an image, we add these examples to a
retraining set as false positives, and do not save the results for human identification. If less than 10 ‘positives’
appear in the images, we then rotate each positive 3 times (in multiples of 90 degrees) and re-predict on that given
tile (image subsection). If the average across all rotations falls below the confidence threshold, that tile is added to a
retraining set as a false positive. If the average confidence across all rotations for a particular tile remains above the
confidence threshold, then it is passed onto the user for human inspection. When we process one flight’s images
(~1,300 in total) we randomly select 5 % before retraining on the false positives, giving them priority amongst the
training set as a whole. We then use the retrained model to predict on the remaining images. If any images were
saturated with false positives, and their results not saved for human review, the program retrains again and uses the
updated model to process the final images.
Results and Discussion: Initial, though limited, tests show a 5-10 fold reduction in the number of false positives
without any noticeable degradation in the algorithm’s ability to recognize meteorite examples presented to it in the
validation set.
Strewn Field Searching. This utility of searching for photographically observed meteorite falls has also recently
been applied to strewn fields that have been uncovered with relic data from satellite sensors and weather radar
stations [6, 7]. By searching a newly identified strewn field, the drone searching methodology led to the collection of
4 additional meteorites (pictured below, yellow boxes are 30x30 cm).
Future Work: The priority for this project is to develop a front-end software package to enable meteorite
searchers without programming experience to easily access and use this software. We will also apply this searching
methodology to our backlog of observed meteorite falls.
References: [1] Devillepoix et al. (2020) Planetary and Space Science 191:105036 [2] Anderson S. L. et al.
(2020) Meteoritics and Planetary Science 55(11):2461-2471 [3] Anderson S. L. et al. (2022) The Astrophysical
Journal Letters 930(2):L25 [4] Hill P. J. et al. (2023) Meteoritics and Planetary Science 58(3):421-432 [5] Citron R.
I. et al. (2021) Meteoritics and Planetary Science 56(6):1073-1085 [6] Fries M. and Fries J. (2010) Meteoritics and
Planetary Science 45(9):1476-1487 [7] Devillepoix H. A. R. et al. (2022) Lunar and Planetary Science Conference
Abstract #2678
