An Introduction to Geospatial Foundation Model

The increasing global concerns regarding climate change and environmental sustainability highlight the urgent necessity for effective solutions. The development of a foundational model for analyzing geospatial satellite images emerges as one such solution. We are currently engaged in the collaborative creation of an open-source geospatial foundational model with NASA. This endeavor entails pre-training large transformer models using multispectral remote-sensing image datasets sourced from diverse channels. Our foundational model demonstrates state-of-the-art performance, requiring fewer labeled training examples and exhibiting robust generalization capabilities across geographical regions. During the session, we will present recent advancements in the geospatial foundational model initiative, including techniques for global region data sampling, the latest experiments in fusing multisatellite images, and downstream applications such as flood detection in Japan.