Global morphospectral map of Mercury through unsupervised learning

Introduction
Geological mapping of planetary surfaces is fundamental to understanding formation history, surface processes, and compositional variations ^1,2, and is critical for landing site selection, mission planning, and exploration support ². Recent efforts have yielded near-complete regional geological maps of Mercury e.g., ^3-6 in support of the ESA-JAXA BepiColombo mission^7,8. However, multi-mapper strategies introduce inconsistencies due to individual interpretation and subjective judgment ⁹, complicating global integration. Additionally, planetary mapping typically relies on photo-interpretative approaches focused on surface morphology, often treating spectral data as supplementary¹⁰. Automated approaches on Mercury have applied deep learning for feature detection ^11-12. While useful for feature catalogues, these methods provide limited geological context regarding terrain character, lithology, or ejecta distribution and does not consent the creation of comprehensive geological maps. To overcome these limits, an automatic morphospectral mapping approach was proposed by Vergara Sassarini et al. (2025) that relies on an unsupervised learning technique based on Gaussian Mixtures which allows obtaining comprehensive explorative maps that merge morphological and spectral information in a single product.
The present study aims to generate the first global morphospectral classification map of Mercury using unsupervised GMM clustering (following¹³), enhanced by a Variational Graph Attention Autoencoder (VGATs¹⁴) preprocessing step to include spatial context and reduce dimensionality. This approach integrates spectral and morphological properties to partition terrains aiming to support the interpretation of unmapped or poorly understood regions, and to evaluate existing maps in the context of Mercury's global mapping effort¹. Furthermore, the developed methodological framework will provide new analytical pipelines and techniques to the higher-resolution data expected from BepiColombo^7,8,15

Methods
The dataset comprises MESSENGER's global DTM and the global MDIS-WAC 8-color mosaic. The methodology follows the unsupervised clustering framework of ¹³, validated on the H05 Hokusai quadrangle and now extended globally. This approach integrates morphological and spectral data through Gaussian Mixture Models (GMM) applied to a multi-dimensional morphospectral datacube. As an improvement over this original framework, an advanced preprocessing method based on a Variational Graph Attention Autoencoders (VGATs¹⁴) will enhance feature selection by performing a dimensionality reduction from a high-dimensional (e.g., 30-dimensional dataset) to a highly informative 3-dimensional latent space. This latent space, which is intrinsically informed by the spatial context of the pixels, provides a highly structured foundation for the subsequent clustering analysis. This spatial regularization effectively prevents the clustering algorithm from fragmenting continuous geological formations, resulting in a much cleaner, more reliable, and physically interpretable geological map. VGATs + GMM clustering will ultimately provide a first version of the global morphospectral map.

Figure 1. Morphospectral classification of the Rachmaninoff basin (H05 Hokusai quadrangle) derived using the GMM-based unsupervised clustering method¹³ shown over the MESSENGER MDIS BDR global basemap in Robinson projection. This regional classification provides the validated foundation for the global morphospectral mapping presented in this study.

Expected results and future developments
The generation of a global morphospectral map of Mercury through unsupervised VGATs + GMM clustering represents a significant step toward automated, data-driven geological interpretation. This map is expected to provide a consistent, planet-wide classification that integrates both morphological and spectral information, capturing regions of compositional or geomorphologic complexity, guiding targeted analyses and prioritizing areas for higher-resolution observations. The resulting framework is directly transferable to other airless bodies with comparable datasets, such as the Moon, expanding its utility beyond Mercury. This capability is particularly relevant in the context of the BepiColombo mission, which will deliver a new generation of high-quality datasets for Mercury⁷. Specifically, the SIMBIO-SYS instrument suite¹⁵ will provide global DTM coverage at significantly higher spatial resolution (50 to 120 m/pixel) and vertical accuracy than MESSENGER through its Stereo Channel (STC), while the VIHI hyperspectral channel will offer coverage at much higher spectral resolution across the visible-near-infrared domain (400-2000 nm). The methodological approach presented here provides a scalable framework for integrating these future BepiColombo datasets into coherent, data-driven products, enabling a more comprehensive understanding of Mercury's surface evolution.

Acknowledgments
This research is funded from the Italian Space Agency (ASI) under ASI-INAF agreement 2024-18-HH.0.

References
1. Galluzzi V. Multi-mapper Projects: Collaborative Mercury Mapping. In: Hargitai H, editor. Planetary Cartography and GIS. Springer International Publishing; 2019. p. 207–18.1.
2. Luna JW, et al. Planetary Geologic Maps: Essential Tools for Scientific Inquiry and Space Exploration. J Geophys Res Planets. 2024;129(10):e2024JE008442.
3. Galluzzi V, et al. Geology of the Victoria quadrangle (H02), Mercury. J Maps. 2016;12(sup1):227–38.
4. Wright J, et al. Geology of the Hokusai quadrangle (H05), Mercury. J Maps. 2019;15(2):509–20.
5. Giacomini L, et al. Geology of the Kuiper quadrangle (H06), Mercury. J Maps. 2022;18(2):246–57.
6. Giacomini L, et al. Geology of Tolstoj quadrangle (H08), Mercury. J Maps. 2024;20(1).
7. Benkhoff J, et al. BepiColombo—Mission Overview and Science Goals. Space Sci Rev. 2021;217(8):90.
8. Rothery DA, et al. Rationale for BepiColombo studies of Mercury's surface and composition. Space Sci Rev. 2020;216(4):66.
9. Rothery D, et al. Mercury's surface and composition to be studied by BepiColombo. Planet Space Sci. 2010;58(1–2):21–39.
10. Nass A, et al. 50 Years of Sensor-Based Planetary Cartography: Review and Perspectives. Proc ICA. 2021;4:1–8.
11. La Grassa R, et al. From the Moon to Mercury: Release of Global Crater Catalogs Using Multimodal Deep Learning for Crater Detection and Morphometric Analysis. Remote Sens. 2025;17(19):3287.
12. Deutsch AN, et al. Hollows on Mercury: Global Classification of Degradation States and Insight Into Hollow Evolution. J Geophys Res Planets. 2025;130(2):e2024JE008747.
13. Vergara Sassarini NA, et al. Explorative geological maps through unsupervised learning. EGU Gen Assem Conf Abstr. 2025;EGU25-19408.
14. Spina F, et al. Variational Graph Attention Autoencoders for dimensionality reduction in morphospectral mapping of planetary surfaces. In: Atti del XXI Congresso Nazionale di Scienze Planetarie. Modena, Italy; 2026.
15. Cremonese G, et al. SIMBIO-SYS: Scientific Cameras and Spectrometer for the BepiColombo Mission. Space Sci Rev. 2020;216(5):75.