A new open-source benchmark wants to do for lunar science what ImageNet did for computer vision.
Researchers introduced Moonstone, a multimodal foundation model and benchmark built from 28 channels of global lunar remote sensing data spanning optical imagery, spectroscopy, thermal emissions, radar, gravity, and elemental composition. The dataset draws from seven instrument families across five orbital missions and covers the entire Moon at roughly 237 meters per pixel. Alongside the dataset, the team released MG-MAE, a masked autoencoder architecture designed to handle the Moon's particular challenges: missing data modalities, uneven spatial coverage, and the need for physically plausible reconstructions across the electromagnetic spectrum.
The benchmark matters because no equivalent has existed before. Decades of lunar orbital data have piled up across disconnected archives, making it nearly impossible to train or evaluate machine learning models in a systematic way. MG-MAE pretrained features outperformed both ImageNet-pretrained baselines and vanilla masked autoencoders across all six downstream tasks — classification, regression, and segmentation — by what the researchers describe as large margins.
The release arrives as space agencies and commercial operators plan a return to the Moon in earnest, with mapping and site-selection decisions increasingly leaning on automated analysis. Whether Moonstone becomes the community standard depends on whether researchers outside the original team adopt it — but with data and code posted publicly on Hugging Face and GitHub, the barrier to entry is low.