Thread by @gencersumbul

🎉 Meet SMARTIES: an open-source, multi-sensor foundation model that lets Earth-observation (EO) data from optical, SAR, VHR RGB and even future sensors speak the same language. 🧵

🌐 Project page: gsumbul.github.io/SMARTIES/
📄 arXiv preprint: arxiv.org/abs/2506.19585
🤗 Model weights: huggingface.co/collections/gs…
👇

Built to unlock processing of multi-sensor EO data in a sensor-agnostic way at global scale, SMARTIES projects heterogeneous EO sensors into a shared spectrum-aware space while staying lightweight enough for your laptop — and powerful enough to top 10+ public benchmarks. 👇

🚀 Why you’ll love it:

🛰️ One model, diverse sensors: optical, SAR, sub-meter RGB, plus unseen sensors in zero-shot mode.
🌈 Spectrum-aware projections: heterogeneous bands are mapped into a single latent space based on frequencies, no complex architecture hacks required.
👇

⚡Lightweight, scalable: available as ViT-B and ViT-L checkpoints with negligible increase in parameter counts, plus fine-tuned/linear probed models.
🔀 Flexible band combinations: use arbitrary combinations of spectral bands for downstream, enabling flexible EO applications.👇

🔄 Powerful downstream transfer: SOA performance using a unified model across a diverse set of EO sensors and tasks beating specialist models.

🎤 See you at #LPS25!

SMARTIES will be presented at ESA’s Living Planet Symposium tomorrow: “A Unified Foundation Model for Multi-Sensor Earth Observation: Breaking Representation Barriers across Remote Sensing Sensors”, 11:30, Hall G1 — drop by if you’re around! 👇

🙌 Huge thanks to my brilliant co-authors Chang Xu, @emasasso & @devistuia — and to the @esa for their support!

#RemoteSensing #EarthObservation #FoundationModel #GeospatialAI #MultiSensor #MultiModal