SomBench: Benchmark Dataset for Advancing Machine Learning in Lunar Science

Himanshu Patil; Gabby Nyirjesy; Rachel A. Slank; Vishal Gaur; Daniela Szwarcman; Paolo Fraccaro; Nikolaos Dionelis; Michael K. Barker

cs.CVSep 8, 2026

SomBench: Benchmark Dataset for Advancing Machine Learning in Lunar Science

Authors: Himanshu Patil, Gabby Nyirjesy, Rachel A. Slank, Vishal Gaur, Daniela Szwarcman, Paolo Fraccaro, Nikolaos Dionelis, Michael K. Barker, +14 more

Organizations: The University of Alabama in Huntsville, Huntsville, AL, USA · IBM Research, Yorktown Heights, NY, USA · Science and Technology Institute, Universities Space Research Association (USRA), Huntsville, AL, USA · NASA Goddard Space Flight Center, Greenbelt, MD, USA · SETI Institute, Mountain View, CA, USA · NASA Ames Research Center, Moffett Field, CA, USA · University of Maryland, Baltimore County (UMBC), Baltimore, MD, USA · Howard University, Washington DC, USA · NASA Headquarters, Washington, DC, USA · NASA Marshall Space Flight Center, Huntsville, AL, USA

Abstract

Lunar orbital missions, such as Lunar Reconnaissance Orbiter, Kaguya/SELENE, Gravity Recovery and Interior Laboratory, and Lunar Prospector, among others, provide rich multi-instrument observations, but their heterogeneity in sampling, projection, and conventions limits reproducible machine learning (ML). We introduce SomBench, a unified, spatially-aligned, ML-ready lunar dataset aggregating 30+ co-registered layers from ten instruments across four missions, spanning 1 meter to 20 kilometer/pixel and covering 82 degree latitude in 90 Lunar Transverse Mercator zones with two polar stereographic caps. An image-anchored tiling pipeline yields pretraining-ready multimodal tile views with leakage-safe splits, distributed as netCDF with Parquet catalogs. An application benchmark suite spans impact processes, volcanic history, and polar volatiles. Baseline experiments with ResNet-50 and SwinV2-B models confirm that each benchmark task is learnable from the released inputs, establishing reference points for future model development.

Figures & tables

Instrument	Mission	Resolution	Bands / Products	Data Type
LROC NAC	LRO	${\sim}1$ m	Panchromatic ( $400-760$ nm)	Individual images, stereo DTMs + related topographic maps
LROC WAC	LRO	100 m – 500 m	VIS: 415, 566, 604, 643, 689 nm; UV: 321, 360 nm; morphologic mosaic; multi-band normalized reflectance mosaics; TiO 2 abundance	Individual images, global image mosaics
LOLA	LRO	20 m – 1 km	Elevation, slope, aspect, roughness, polar 1064 nm normal albedo, polar illumination, PSRs	Lidar-based global and polar DTMs + related topographic maps; other derived products
Diviner	LRO	240 m – 15 km	$0.35-400\mu\rm m$ ; rock abundance, $T_{\mathrm{reg}}$ anomaly, H-parameter, $T_{\mathrm{bol}}$ (24 phases), ice stability depth	Thermal emission derived global and polar maps
Mini-RF	LRO	90 m	Monostatic S-band radar (12.6 cm) CPR, S1 reflectivity	Controlled, orthorectified, geometrically-normalized global and polar mosaics
TC	Kaguya/SELENE	60 m	Panchromatic ( $430-850$ nm); elevation, slope, aspect	Global stereo DTM mosaic + related topographic maps

Table 1: Missions, instruments, and data products included in SomBench.

Region ID	NAC PHO Site Name	Number of images
E009S3481	Apollo 15 SIVB Impact	83
E010N0230	Apollo 11	99
E011N1499	Highlands	88
E018N3346	Mare Impact Melt	54
E041N1229	King Crater Ejecta	32
E074N3010	Reiner Gamma	20

Table 2: Co-registered NAC Photometry (PHO) locations and number of images.

Figure 1: The distribution of NAC, NAC Feature Specific, NAC IMP, NAC PHO, NAC DTMs and DTM Orthomosaic images over the lunar surface in the SomBench dataset.

Figure 2: Image presents the distribution of WAC, WAC Feature Specific and WAC PHO Site images used in the SomBench Dataset.

#	Dataset	Instrument / Source	Global (res.; coverage)	Polar (res.; lat., both poles)	Key Purpose / Products	Ref.
1	LROC WAC Global Morphologic Mosaic	LRO WAC (643 nm)	100 m/px; 90°S–90°N	100 m/px; 60–90°	Surface morphology emphasis: crater rims, wrinkle ridges, flow fronts, blocky ejecta, topographic shading	[ 31 , 32 , 33 , 34 ]
2	LROC WAC Normalized Multi-band Reflectance Mosaics	LRO WAC (7-band UV–VIS)	500 m/px; 60°S–60°N	—	Photometrically corrected reflectance (I/F); comparative spectral & compositional analysis; band-ratio work	[ 35 , 36 , 37 ]
3	LROC WAC Normalized Higher-Resolution Reflectance Mosaic	LRO WAC (643 nm)	100 m/px; 90°S–90°N	100 m/px; 60–90°	Higher-resolution morphology / registration anchor for the WAC normalized reflectance suite	[ 35 , 37 , 33 , 34 ]
4	Kaguya MI Multi-band Normalized Reflectance Mosaics	Kaguya MI (8-band VIS–NIR)	60 m/px; 55°S–55°N	—	Mafic mineral absorption ( $\sim$ 1 µm); broadband maturity & albedo; spectral comparison across terrains	[ 5 , 23 , 6 , 7 ]
5	Kaguya MI Derived Mineralogy	Derived from MI multi-band reflectance inversion	60 m/px, 1 km/px; 55°S–55°N	—	60 m/px: olivine, OPX, CPX, plagioclase, FeO, plag. grain size, OMAT. 1 km/px: smFe, mpFe, npFe	[ 5 , 6 , 7 , 38 ]
6	Kaguya SP Derived Mineralogy	Derived from Kaguya SP hyperspectral inversion	—	1 km/px; 80–90°	Olivine, OPX (low-Ca), CPX (high-Ca), plagioclase, FeO, OMAT, npFe	[ 24 , 25 ]

Table 3: Summary of lunar global static map products. A dash indicates a product released for only one of the two regions; polar latitude ranges apply at both poles.

Figure 3: Tile extraction pipeline for the image-anchored low-resolution and high-resolution tracks.

Panel of Figure 4 — Figure 4: Example of low-resolution non-polar ( top ) and polar ( bottom ) samples, displaying all modalities side by side. Plots include the first channel of each modality at 51.2 x 51.2 km resolution, or 512 x 512 WAC visible pixels. Certain modalities, such as PSR and ALBEDO , only exist at the poles and are thus missing from the non-polar sample. Similarly, other modalities, such as TIO2 and ROCK_ABUNDANCE , do not exist at the poles.

Panel of Figure 5 — Figure 5: Example of high-resolution non-polar ( top ) and polar ( bottom ) samples, displaying all modalities side by side. Plots include the first channel of each modality tile at 512 x 512 m resolution, or 512 x 512 NAC pixels. Certain modalities, such as PSR and ALBEDO only exist at the poles and are therefore missing from the non-polar sample. Similarly, other modalities, such as TIO2 and ROCK_ABUNDANCE do not exist at the poles.

Component	Low-Resolution Track	High-Resolution Track
Optical Anchor	WAC VIS (100 m/px)	NAC (1 m/px)
Tile Resolution	51.2 km $\times$ 51.2 km	512 m $\times$ 512 m
Tile Size (pixels)	512 $\times$ 512	512 $\times$ 512
Rejection stride	120	120
Pass 1 Extracts	VIS, UV, DTM (60m)	NAC only
Pass 2 Extracts	Static maps	NAC-derived DTMs + Static maps

Table 4: Comparison of low-resolution and high-resolution track characteristics.

Modality	Dataset key	Native Resolution (m)	Low-resolution count	Low-resolution avg. null (%)	Low-resolution snapped resolution (m)	High-resolution count	High-resolution avg. null (%)	High-resolution snapped resolution (m)
NAC Panchromatic	NAC	1.0	–	–	–	1,000,113	0.14%	1.0
WAC VIS ( $415\text{\,}\mathrm{nm}689\text{\,}\mathrm{nm}$ )	WAC_VIS	100.0	963,609	0.24%	100.0	–	–	–
WAC UV ( $321\text{\,}\mathrm{nm}360\text{\,}\mathrm{nm}$ )	WAC_UV	500.0	963,609	0.26%	501.96	–	–	–
Topography (SLDEM2015)	DTM_60M (high-res) DTM (low-res)	60.0	963,609	0.00%	60.02	1,000,113	0.00%	56.89
NAC-stereo DTM	DTM_3M	3.0	–	–	–	999,824	0.12%	2.99
Topography (SLDEM2015) Slope	SLOPE_60M (high-res) SLOPE (low-res)	60.0	963,609	0.00%	60.02	1,000,113	0.00%	56.89

Table 5: Per-modality coverage for the low-resolution and high-resolution tracks. Counts indicate the number of emitted netCDF files for each modality, and Avg Null Pct is the average percent of null pixels per tile. Dataset key indicates the parquet column associated with this modality. Modality resolution for both the high-resolution and low-resolution datasets after resampling procedure also included.

Split	Images	Annotations	Craters / img	Median $\sqrt{\text{area}}$	Incidence (°)	Emission (°)
Train	800	91,516	114.4	16.5	72.2	1.81
Val	100	11,429	114.3	16.5	71.1	1.55
Test	100	11,384	113.8	17.0	72.7	1.39

Table 6: WAC Robbins COCO dataset splits with per-split image and annotation counts, mean craters per image, median crater size (

\sqrt{\text{area}}

in px), and mean illumination geometry.

Figure 6: Hand-labeled crater annotations on two NAC study areas, Apollo 15 S-IVB (left) and Apollo 17 (right). Green boxes denote annotated crater instances. Apollo 15 S-IVB was labeled at 5 m/pixel resolution, causing it to appear blurrier than the Apollo 17 image which was labeled at a higher resolution.

Name	Science Theme	Image Size (px)	Resolution (m/px)	Target	Train	Val	Test
Object Detection
Robbins craters	Impact processes	512 $\times$ 512	100	1 class	800	100	100
NAC hand labeled	Impact processes	256 $\times$ 256	0.8–5	1 class	289	56	63
Segmentation
Irregular Mare Patches	Volcanic History	256 $\times$ 256	1	1 class	100	20	10
Regression

Table 7: Overview of SomBench benchmark datasets separated by task type.

Model	mAP	AP@50	AP@75
ResNet-50 (ImageNet)	0.1411 $\pm$ 0.0036	0.4374 $\pm$ 0.0196	0.0895 $\pm$ 0.0055
SwinV2-B (ImageNet)	0.1552 $\pm$ 0.0086	0.4586 $\pm$ 0.0308	0.1090 $\pm$ 0.0146

Table 8: Crater detection on the NAC hand-labeled benchmark (meter-scale LROC NAC imagery). Test-set bounding-box mean average precision: mAP (IoU 0.50:0.95), AP@50, and AP@75. Values are mean

\pm

standard deviation over five random seeds.

Figure 7: Qualitative crater detection results on three test tiles of the NAC hand-labeled benchmark. Columns show the input tile and predictions from SwinV2-B, and ResNet-50 (ImageNet), using the released weights. Green boxes denote ground-truth annotations and red boxes model predictions with confidence $\geq$ 0.5. Per-panel insets report F1@0.5 and true-positive, false-positive, and false-negative counts. — Figure 7: Qualitative crater detection results on three test tiles of the NAC hand-labeled benchmark. Columns show the input tile and predictions from SwinV2-B, and ResNet-50 (ImageNet), using the released weights. Green boxes denote ground-truth annotations and red boxes model predictions with confidence $\geq$ 0.5. Per-panel insets report [email protected] and true-positive, false-positive, and false-negative counts.

	50% training data			100% training data
Model	mAP	AP@50	AP@75	mAP	AP@50	AP@75
ResNet-50 (ImageNet)	0.1993 $\pm$ 0.0007	0.4960 $\pm$ 0.0036	0.1572 $\pm$ 0.0038	0.2148 $\pm$ 0.0018	0.5288 $\pm$ 0.0055	0.1762 $\pm$ 0.0038
SwinV2-B (ImageNet)	0.2313 $\pm$ 0.0027	0.5849 $\pm$ 0.0072	0.1862 $\pm$ 0.0041	0.2420 $\pm$ 0.0047	0.6020 $\pm$ 0.0061	0.2028 $\pm$ 0.0086

Table 9: Crater detection on the WAC Robbins catalog benchmark (context-scale LROC WAC imagery) as a function of training-data fraction (left: 50% training data, right: 100%). Test-set bounding-box mean average precision: mAP (IoU 0.50:0.95), AP@50, and AP@75. Values are mean

\pm

standard deviation over five random seeds.

Figure 8: Qualitative crater detection results on three test tiles of the WAC Robbins catalog benchmark at 100% training data. Columns show the input tile and predictions from ResNet-50 (ImageNet) and SwinV2-B, using the released weights. Green boxes denote ground-truth annotations and red boxes model predictions with confidence $\geq$ 0.5. Per-panel insets report F1@0.5 and true-positive, false-positive, and false-negative counts. — Figure 8: Qualitative crater detection results on three test tiles of the WAC Robbins catalog benchmark at 100% training data. Columns show the input tile and predictions from ResNet-50 (ImageNet) and SwinV2-B, using the released weights. Green boxes denote ground-truth annotations and red boxes model predictions with confidence $\geq$ 0.5. Per-panel insets report [email protected] and true-positive, false-positive, and false-negative counts.

Model	IoU 1	F1 1
ResNet-50 (ImageNet)	0.4731 $\pm$ 0.0114	0.6423 $\pm$ 0.0105
SwinV2-B (ImageNet)	0.5555 $\pm$ 0.0159	0.7141 $\pm$ 0.0132

Table 10: Irregular Mare Patch (IMP) segmentation. Since IMP delineation is a binary dense-prediction task, we report test-set metrics on the positive (IMP) class only: intersection-over-union (IoU 1 ) and F1 score (F1 1 ). Values are mean

\pm

standard deviation over five random seeds.

Figure 9: Qualitative IMP segmentation results on three test tiles. Columns show the input tile, the ground-truth mask, and predictions from ResNet-50 (ImageNet) and SwinV2-B, using the released weights. Yellow denotes the positive (IMP) class, and per-panel insets report the per-tile IoU.

Model	RMSE $\downarrow$	MAE $\downarrow$	$R^{2}$ $\uparrow$
ResNet-50 (ImageNet)	0.0788 $\pm$ 0.0004	0.0568 $\pm$ 0.0004	0.9162 $\pm$ 0.0009
SwinV2-B (ImageNet)	0.0377 $\pm$ 0.0004	0.0256 $\pm$ 0.0003	0.9808 $\pm$ 0.0004

Table 11: Polar ice prospectivity regression on the eight-modality polar stack (aspect, slope, ice stability depth, maximum temperature, PSRs, PSR density, distance to PSRs, curvature). Test-set metrics against the prospectivity map in [ 65 ] : for RMSE and MAE lower-is-better, and for

R^{2}

higher-is-better. Values are mean

\pm

standard deviation over five random seeds.

Figure 10: Qualitative ice-prospectivity regression results on three polar test patches. Columns show a representative input layer, the target prospectivity map, and predictions from ResNet-50 (ImageNet) and SwinV2-B, using the released weights; per-panel insets report the per-patch MAE.

Explore similar work

Moonstone: A Multimodal Foundation Model and Benchmark for Lunar Remote Sensing
Jul 3, 2026Ayush Prasad, Swarnalee MazumderMissing-Modality Learning Masked Autoencoders
Space and Earth Observation Centre, Finnish Meteorological Institute, Finland · Data Management, German Climate Computing Centre, Germany
MoonAnything: A Vision Benchmark with Large-Scale Lunar Supervised Data
Apr 1, 2026Clémentine Grethen, Yuang Shi, Simone Gasparini +1Remote Sensing Benchmarks 3D Reconstruction
IRIT - Université de Toulouse, France · National University of Singapore, Singapore · IPAL, IRL2955, Singapore
Multimodal-Multiresolution Foundation Model for Lunar Remote Sensing
Sep 8, 2026Paolo Fraccaro, Gabby Nyirjesy, Daniela Szwarcman +19Multimodal Foundation Models Remote Sensing
IBM Research, Yorktown Heights, NY, USA · The University of Alabama in Huntsville, Huntsville, AL, USA · University Space Research Association (USRA), Science and Technology Institute, Huntsville, AL, USA +7

Jul 3, 2026J/K move · Enter open · S save

Moonstone: A Multimodal Foundation Model and Benchmark for Lunar Remote Sensing

Ayush Prasad, Swarnalee Mazumder

Space and Earth Observation Centre, Finnish Meteorological Institute, Finland · Data Management, German Climate Computing Centre, Germany

Missing-Modality Learning Masked Autoencoders Multimodal Pretraining Remote Sensing Remote Sensing Benchmarks

Open paper AskPDF arXiv

Decades of orbital missions have produced multi-modal remote sensing data for the Moon, spanning optical imagery, spectroscopy, thermal emission, radar, gravity, and elemental composition. Yet these datasets remain fragmented across archives, and no benchmark exists for evaluating machine learning on lunar data. We introduce Moonstone, the first multi-modal foundation model benchmark for lunar remote sensing. Our contributions are: (1) a 28-channel, 128 pixels-per-degree (~237 m) global lunar pretraining dataset from seven instrument families across five missions, (2) MG-MAE, a modality-grouped masked autoencoder with per-group convolutional tokenizers, a shared Vision Transformer encoder, attention masking for missing modalities, coverage-adaptive masking for heterogeneous spatial coverage, and spectral continuity regularization for physically plausible reconstructions, and (3) a benchmark of six downstream tasks covering classification, regression, and segmentation. MG-MAE pretrained features outperform scratch baselines on all tasks and surpass both ImageNet-pretrained and vanilla MAE baselines by large margins. Data and code are available at https://huggingface.co/datasets/ayushprd/Moonstone and https://github.com/ayushprd/Moonstone .

Abstract

Explore similar work

Moonstone: A Multimodal Foundation Model and Benchmark for Lunar Remote Sensing

MoonAnything: A Vision Benchmark with Large-Scale Lunar Supervised Data

Multimodal-Multiresolution Foundation Model for Lunar Remote Sensing