EarthMind, a cutting-edge foundational AI model designed to advance our understanding of Earth Observation (EO) satellite data. Developed in collaboration with the University of Trento, TU Berlin, Technical University of Munich, and the University of Pisa, EarthMind represents a significant step forward in the democratization of access to space data.
EarthMind is a multimodal AI model capable of answering complex questions using visual, radar, and multi-spectral data—enabling insights even through clouds and in low-light conditions. It achieves state-of-the-art performance on EarthMind-Bench, a new benchmark introduced alongside the model, featuring over 2,000 human-annotated multi-sensor image-question pairs. Notably, EarthMind outperforms GPT-4o despite its compact 4B parameter size.
Meet the team behind EarthMind: Yan Shu, Bin Ren, Zhitong Xiong, Danda Pani Paudel, Luc Van Gool, Begum Demir, Nicu Sebe, and Paolo Rota.
This research is part of INSAIT’s mission to harness AI for broad societal benefits and aligns with the upcoming GAIA Symposium on Geospatial AI and Applications with Foundation Models, which we are proud to host in Sofia, Bulgaria, from 24–26 September 2025. Registrations are now open – we invite you to join us in shaping the future of Space and AI.
Learn more and register here: gaia.insait.ai/registration
Explore EarthMind:
Preprint: https://arxiv.org/pdf/2506.01667
Software: https://github.com/shuyansy/EarthMind
Model: https://huggingface.co/sy1998/EarthMind-4B
Training Data: https://huggingface.co/datasets/sy1998/EarthMind-data 🤗 Benchmarks: https://huggingface.co/datasets/sy1998/EarthMind-Bench