Democratizing Foundation Models for Earth Observation Applications

A. Zappacosta, R. S. Kuzu, C. O. Dumitru, D. E. Molina, T. Brunschwiler, R. Kienzler, J. Jakubik, B. Blumenstiel, G. Cavallaro, S. Kesselheim, R. Sedona, A. Wijata, L. Tulczyiew, D. Marek, J. Nalepa
‍

In Earth Observation (EO), integrating Foundation Models (FMs) poses a challenge due to disparities with conventional computer vision tasks, particularly in processing multispectral/hyperspectral data. Additionally, the high computational resources needed for FM training limit their widespread use. Leveraging state-of-the-art FMs, we extend their utility to suit diverse sensor types, temporal characteristics, and domains, enhancing their applicability in various EO tasks. Our approach offers a flexible solution, accommodating different multi-modal FMs, and extends beyond specific models for broader EO applications. To demonstrate the effectiveness of this framework, we explore diverse EO use cases (UCs), spanning different thematic areas and geographic locations, and showcase the framework's potential for generalization and adaptability across a range of EO challenges, as part of the FAST-EO project.

‍

‍

‍