Earth Embeddings: Towards AI-centric Representations of our Planet

Konstantin Klemmer, Esther Rolf, Marc Russwurm, Gustau Camps-Valls, Mikolaj Czerkawski, Stefano Ermon, Alistair Francis, Nathan Jacobs, Hannah Rae Kerner, Lester Mackey, Gengchen Mai, Oisin Mac Aodha, Markus Reichstein, Caleb Robinson, David Rolnick, Evan Shelhamer, Vincent Sitzmann, Devis Tuia, Xiaoxiang Zhu

This paper presents a new perspective for the flexible and efficient representation of geospatial data, tailored to and empowered by AI: Earth embeddings. Earth embeddings provide a unified and accessible vector representation of local geographic characteristics. They fuse different geospatial data sources across time and space, compress highly-correlated raw geospatial data into one dense representation, can be used to guide interpolation between data observations, and can serve as a universal location token for foundation models. This provides a powerful alternative to existing geospatial workflows that rely on heterogeneous data, hard-to-acquire expertise, and significant computation by the user: embeddings instead provide convenient representations, easily adaptable for numerous downstream tasks. We posit that Earth embeddings redefine geospatial analytics, transforming it from fragmented, task-specific modeling into a coherent, generalizable framework for AI. We approach this from both the users' and developers' perspectives, outlining a path for how the rapidly developing technology of Earth embeddings can reshape the way we store, represent, and use geospatial data, evidenced by recent research charting initial directions. We call on Earth embedding users and developers to align methodological and applied development and deployment within an interdisciplinary, open-source oriented research community.