---
license: cc-by-4.0
task_categories:
- image-classification
- image-segmentation
tags:
- lunar
- moon
- remote-sensing
- multimodal
- foundation-model
- planetary-science
pretty_name: Moonstone
size_categories:
- 10K<n<100K
configs:
- config_name: benchmark
  data_files: benchmark/*
- config_name: pretraining
  data_files: pretraining/mmap/*
---

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

28-channel, 128 pixels-per-degree (~237 m/pixel) global multimodal lunar dataset assembled from
seven instrument families across five missions (LRO WAC/LOLA/Diviner/Mini-RF, Chandrayaan-1 M3,
GRAIL, Lunar Prospector GRS, Clementine). All channels are aligned to a common equirectangular
grid (46,080 x 23,040 px, lunar sphere a=b=1,737,400 m) and organized into 7 physical modality
groups (surface, thermal, spectral_M3, gravity, radar, hapke, composition).

## Repository layout

The dataset is split into two independent parts so you can download only what you need.

### `pretraining/` — self-supervised pretraining data

| Path | Description |
|------|-------------|
| `pretraining/aligned/` | 28 source-of-truth instrument GeoTIFFs at 128 ppd (+ M3 geometry) |
| `pretraining/mmap/` | Pre-normalized (z-scored) memory-mapped float32 arrays for pretraining (unlimited random crops) + NaN masks + `channel_index.json` |
| `pretraining/channel_stats.json` | Per-channel (mean, std) normalization statistics (200 random 256x256 windows) |

### `benchmark/` — downstream evaluation data

| Path | Description |
|------|-------------|
| `benchmark/lunar_patches_v4.h5` | 16,200 patches (180x90 grid) x 28 x 256 x 256, with geology/age/mare metadata and fixed 70/15/15 split |
| `benchmark/geologic_units.tif`, `benchmark/geologic_units.json` | USGS geologic-unit label raster + class map (geology + age tasks) |
| `benchmark/mare_mask.tif` | Mare vs. highlands ground truth (segmentation) |
| `benchmark/crater_mask.tif` | Crater (over 10 km) ground truth (segmentation) |

## Downloading

Download just one part (each is independent):

```python
from huggingface_hub import snapshot_download

# Benchmark only (evaluation data + labels, ~74 GB)
snapshot_download("ayushprd/Moonstone", repo_type="dataset",
                  allow_patterns="benchmark/*", local_dir="Moonstone")

# Pretraining only (z-scored arrays + source GeoTIFFs, ~207 GB)
snapshot_download("ayushprd/Moonstone", repo_type="dataset",
                  allow_patterns="pretraining/*", local_dir="Moonstone")
```

## Channels (28)

surface: wac_morphology, elevation, slope, roughness · thermal: diviner_tbol_midnight,
diviner_temp_night, rock_abundance, christiansen_feature · spectral_M3: m3_{750,950,1000,1250,1580,2000,2817,2857} ·
gravity: grail_{freeair,bouguer,uncertainty} · radar: minirf_{cpr,s1} (log1p) ·
hapke: wac_hapke_{415,566,604,689}nm · composition: clementine_uvvis_750nm, lpgrs_{tio2,feo}

## Benchmark tasks

![Moonstone benchmark tasks](benchmark_overview.png)

Six downstream tasks define the benchmark.
Geology (49-class), Age (5-class), Composition (FeO and TiO2 regression), Cross-modal thermal
prediction, Mare and highlands segmentation, Crater (over 10 km) segmentation.

## Code

Pretraining and benchmark code (MG-MAE model, the 15-step data pipeline, and downstream
evaluation for all six tasks): <https://github.com/ayushprd/Moonstone>

## Provenance
All data derived from public NASA PDS / USGS / ODE archives. Built via the 15-step pipeline in
the Moonstone code repository (steps 01-15 + fix_minirf). Normalization: z-score, NaN->0 after norm.