Tishampati Dhar: A review of Geospatial Foundation models, Embeddings and Practical Applications
As earth observation sensors proliferate, processing data from all of the sources and assimilating them into earth systems models becomes more and more challenging due to sheer volume and computational needs. Previously, access to space was a limitation and sensor data was scarce, now the bottleneck has moved down to efficient use of this data. The emerging and now relatively established solution to this deluge is capturing the shared latent space being observed by these sensors in form of embeddings, which preserve the fidelity of the observations, but compress the data volumes. A modern deep learning driven progression to the philosophical paths of data fusion research in the past decades. This presentation will focus on the state of the art of Geospatial Foundation models and how they can be integrated into physics driven earth systems models.