Embedding Geography: Spatial Representation Learning for Foundation Models, Aalto University, 14.11.2025
The rise of large-scale pre-trained models, also known as foundation models, has sparked great interest within the geospatial and urban analytics communities. Despite their impressive capabilities, general-purpose foundation models remain limited in performing complex geospatial analyses. This reveals the need for developing dedicated geospatial foundation models. To this end, we view learning effective representations from multi-modal geospatial data as a key pathway toward building such models. In this talk, I will present our progress in learning spatial representations (and thus foundation models) across diverse modalities, such as OpenStreetMap, human trajectories, and remote sensing/street view images. I will conclude by discussing major challenges and promising future directions in geospatial foundation model research.