A large share of globally available data can be geo-referenced; mapped onto the sphere of planet Earth.

These dependencies of geospatial data can be leveraged to train large, unsupervised "foundation" models that fuse different layers of spatial information and learn meaningful representations of every location on the planet.

Such Large Geospatial Models (LGMs) will be transformative for private and public sector alike. From visionary moonshots like planetary digital twins, to high-impact commercial applications in the insurance sector, LGMs promise to be a general-purpose toolbox—similar to Large Language Models (LLMs).

Building LGMs requires tackling several distinct challenges, centered around (1) building dedicated neural network architectures, (2) collecting and maintaining data, (3) computational scaling, (4) constructing training objectives and (5) robust evaluations. With these challenges in mind, we will also briefly assess the geospatial understanding of state-of-the-art LLMS.

Lastly, we will present a prototype LGM and highlight promising results on geospatial downstream tasks.

‍