July 8th, 2026 | 14:00 CEST
Live on Microsoft Teams.
On July 8th, the ESA Φ-Lab Collaborative Innovation Network will host a new Φ-talk. Details are below.
Meet the speakers
Gencer Sumbul is a postdoctoral researcher at the Environmental Computational Science and Earth Observation (ECEO) Laboratory, EPFL, Switzerland. His research focuses on advancing computer vision and machine learning methodologies for multimodal Earth observation (EO), with a current emphasis on sensor-agnostic foundation models. He received his PhD from the Faculty of Electrical Engineering and Computer Science at TU Berlin, where he created BigEarthNet, one of the most widely used multimodal benchmark datasets for EO research. He is also the creator of SMARTIES, a spectrum-aware foundation model for multi-sensor remote sensing. He has authored over fifty publications with more than 2000 citations, and has contributed to several European Space Agency and Horizon Europe research projects aimed at scalable and generalizable AI for environmental and societal understanding.
Talk abstract
Foundation models offer a promising route for extracting transferable representations from large-scale Earth observation (EO) archives. Yet, their application to remote sensing (RS) remains constrained by important limitations: many existing models remain tied to particular sensors or fixed sensor combinations, and use language supervision that is often too shallow to support rich EO reasoning. This talk presents the findings from our ESA project on building a foundation model for multimodal EO. The project investigated how representation learning across diverse RS sensors can be unified through SMARTIES, a spectrum-aware sensor-agnostic foundation model. SMARTIES projects multi-sensor EO data into a shared spectral space and reconstructs masked, mixed-sensor inputs through cross-sensor token exchange within a unified transformer framework, enabling multi-sensor pretraining and downstream transfer without the need for sensor-specific reconfiguration or retraining, significantly improving both efficiency and generalization. Furthermore, the project also investigated how SMARTIES can be bridged to language semantics by leveraging large language models without the need for retraining/finetuning.
Register here!