Global Dataset of Boulders on the Moon from LRO NAC Images using Deep Learning

Global dataset of boulders on the Moon from LRO NAC images using Deep Learning. B. Aussel1, O. Rüsch1,2, B. Gundlach1, V. T. Bickel3, S. Kruk4, E. Sefton-Nash4, 1Universität Münster, GER ([email protected]), 2Space Exploration Institute, CH, 3University of Bern, CH, 4European Space Agency, NL/ES. Introduction: Boulders on the surface of the Moon are typically located within and around fresh impact craters and on steep slopes [1], while the main source of boulders are impacts excavating bedrock [2,3]. Smaller meteoroid impacts limit their survival time on the surface after emplacement, shattering the boulders and leading to regolith production [4-7]. Analysis of the locations and sizes of individual boulders can indicate the age of a surface unit [6,7] and subsurface rock content or regolith thickness [e.g., 8,9]. Currently, the only existing global rock abundance map of the lunar surface was derived by NASA’s Lunar Reconnaissance Orbiter (LRO) Diviner instrument [10], using the contrasting thermophysical properties of rocks and regolith [1,11], which, however, does not contain information about individual boulders. These can be identified in LRO’s Narrow Angle Camera (NAC) [12] images with highest ground-sampling resolutions down to ~0.5 m/px, covering almost the entire lunar surface [e.g., 6,7]. To obtain a global dataset, automatic boulder identification needs to be performed, since manual mapping is infeasible due to the large number of images. The utilization of deep learning-based methods has been proven to allow for a fast and reliable identification of geologic features in planetary surface images, such as rockfalls, craters, and boulders [e.g., 13-17]. Methods: Using the deep learning YOLOv8 architecture in Python [18,19], we developed an automatic boulder detection algorithm. For this, we manually mapped 72,097 boulders on 985 representative NAC image tiles with a bounding box around the illuminated part. This dataset was split into 70% for training, 10% for validation, and 20% for testing of the boulder detector. Average precisions between 82.5% and 88.5% for an Intersection-over-Union (IoU, measure of bounding box accuracy) threshold of 0.5 and 93.3% and 98.4% for an IoU threshold of 0.1 are achieved. We applied the detector to the lunar surface between 60°N and 60°S, by selecting NAC images based on resolution. Complete boulder mapping becomes increasingly challenging above 60°N and below 60°S due to large regions hidden by shadows caused by high incidence angles. The image resolutions were restricted to ≤ 1.5 m/px to use the highest resolution images, and the incidence angles to above 35° and below 75° to ensure that the rocks are casting shadows. The lunar surface was divided into grid cells of size 0.02° and for each cell, from the available covering images, the highest resolution image Figure 1: Examples of the automatically mapped boulders (green boxes) for LRO NAC image M145047737LC. was selected. In total, 635,386 NAC images were analyzed. Results: The derived dataset contains 95,796,361 boulders larger than 4.5 m, with 63,891,702 (67%) boulders located in the lunar highlands and 31,904,659 in the lunar maria (33%, following the selenographic definition by [20]). With increasing diameter, the boulder density, area and aspect ratio decrease. We find a mean diameter of ~6.5 m, area of ~32.8 m2 and aspect ratio between the long and short axis of ~1.31. Boulders can preferentially be found at impact craters (crater walls, interior and ejecta) and steep slopes associated with e.g., wrinkle ridges or rilles, as expected. In addition to the known generally higher boulder density in the maria compared to the highlands, we also find a difference in the mean boulder diameter (Fig. 2), with ~6.8 m in the highlands, compared to ~6.0 m in the maria. This is reflected in the size-frequency distribution showing a substantially higher boulder density in the maria below ~10 m and a steeper slope. Discussion: The observed higher boulder density in the maria agrees with the Diviner rock abundance map [1,11] and can be explained by several factors. Firstly, there might be a contribution by the preferential location of wrinkle ridges in the maria [21,22]. Nevertheless, boulders are predominantly located within and around impact craters [2,3]. Secondly, boulders emplaced on the surface during crater formation might have different physical properties in the maria and highlands (e.g., mechanical strength) that could lead to shorter survival times of highlands boulders [5]. Thirdly, the varying regolith thickness or subsurface rock content between the highlands and maria likely contributes to the differences [e.g., 9]. Craters of similar size approximately have similar excavation depths [2]. Therefore, during the formation of equally sized craters, a thinner regolith layer contributes with a lower fraction to the ejected material compared to competent boulders, originating from bedrock [9]. This leads to a higher ejecta boulder density around the crater. Our analysis of ejected boulders around recently formed impact craters confirms this hypothesis. For craters smaller than 1.5 km, the mean boulder densities between maria and highlands crater ejecta are distinguishable, indicating a thinner regolith layer or higher subsurface rock content in the maria. We confirm this observation by measuring the boulder distributions around fresh “cold spot” craters (Fig. 3) [24]. Outlook: Additional open questions, such as the variability of regolith thickness or mechanical and thermophysical properties of boulders across the lunar surface, can be addressed with our dataset. This can be achieved by combining our boulder inventory with existing datasets such as the Diviner rock abundance map [11] or Mini-RF data [25]. References: [1] Bandfield et al. (2011) JGR, 116. [2] Melosh (1984) Icarus, 59. [3] Hartmann (1969) Icarus, 10. [4] Hörz et al. (2020) PSS, 194. [5] Basilevsky et al. (2013) PSS, 89. [6] Rüsch et al. (2022) Icarus, 387. [7] Rüsch & Aussel (2024) JGR: Planets, 129. [8] Bart & Melosh (2010) JGR, 115. [9] Elder (2019) JGR: Planets, 124. [10] Paige et al. (2010) Space Sci. Rev., 150. [11] Powell et al. (2023) JGR: Planets, 128. [12] Robinson et al. (2010) Space Sci. Rev., 150. [13] Bickel et al. (2019) IEEE TGRS 57. [14] Bickel et al. (2020) Nature Comm., 11. [15] Fairweather et al. (2022) Earth and Space Science, 9. [16] Zhu et al. (2021) Remote Sens., 133. [17] Prieur et al. (2023) JGR: Planets, 128. [18] Redmon et al. (2016) IEEE CVPR 779-788. [19] Jocher et al. (2023) Ultralytics YOLO (Version 8.0.0), Github. [20] Nelson et al. (2014) LPSC 45, Abstract #2861. [21] Yue et al. (2015) JGR: Planets, 120. [22] Thompson et al. (2017) LPSC 48, Abstract #2665. [23] Lemelin et al. (2019) PSS, 165. [24] Williams et al. (2018) JGR: Planets, 123. [25] Cahill et al. (2014) Icarus, 243. Figure 2: Map of the mean boulder diameter for boulders larger than 4.5 m. White areas represent gaps in the map. Figure 3: Mean of the size-frequency distributions for the boulders located around multiple (n) lunar cold spot [24] craters (within 1.2 and 3 crater radii distance to the crater center) in the maria and in the highlands. The gray dashed line indicates the utilized 4.5 m boulder diameter threshold.