2026-10-09, 12:30–13:00 (Europe/Amsterdam), Aula Magna
Satellite Earth Observation (EO) time series are fundamental to monitoring our planet's changing environment. However, inconsistent revisit times and frequent cloud obstruction in optical data (Sentinel-2) often force practitioners to rely on lossy data compositing, which discards critical phenological information.
In this keynote, we introduce TESSERA (Temporal Embeddings of Surface Spectra for Earth Representation and Analysis), a pixel-wise foundation model designed to overcome these challenges. TESSERA leverages multi-modal fusion of Sentinel-1 (radar) and Sentinel-2 (optical) data, employing a self-supervised learning framework based on Barlow Twins and random temporal sampling. This approach ensures high robustness to irregular sampling and missing data without requiring expensive ground-truth labels.
A key highlight of TESSERA is its scale and commitment to Open Science: trained on a global dataset spanning 2017–2025, the model provides high-dimensional temporal embeddings that capture the "spectral fingerprint" of the Earth's surface. In alignment with the FAIR principles, we are committed to making TESSERA an open-access resource for the community. We will demonstrate how TESSERA achieves state-of-the-art performance in downstream tasks such as crop type mapping and land cover classification with minimal labeled data, paving the way for the next generation of open-source, distributed GeoAI monitoring systems.
In this keynote, we will dive deeper into the practical implications of the TESSERA foundation model for the Earth Observation and GeoAI communities. Beyond the architectural innovations presented at CVPR 2026, this session will focus on three key pillars:
- Breaking the "Label Bottleneck": We will discuss how TESSERA’s self-supervised temporal embeddings allow researchers and organizations to build high-performing monitoring tools with 10x to 100x less labeled data than traditional supervised methods.
- A New Paradigm for Multi-Modal Integration: We will showcase how TESSERA natively fuses Sentinel-1 SAR and Sentinel-2 optical data at the pixel level, providing a robust solution for regions with persistent cloud cover.
- Commitment to Open & Distributed Science: In line with the Open-Earth-Monitor mission, we will outline our roadmap for releasing next-gen pre-trained weights and the global embedding dataset (2017–2025).
Zhengpeng (Frank) Feng is a second-year Ph.D. candidate in the Energy and Environment Group, Department of Computer Science and Technology, at the University of Cambridge. His research interests lie at the intersection of machine learning and earth sciences, with a particular focus on developing self-supervised learning methods in remote sensing.