Applications of the TESSERA Geospatial Foundation Model to Diverse Environmental Mapping Tasks
Frank Feng, Clement Atzberger, Sadiq Jaffer, Jovana Knezevic, , Robin Young, Madeline Lisaius, Markus Immitzer, Toby Jackson, James G. C. Ball, David Coomes, Anil Madhavapeddy, Andrew Blake, and Srinivasan Keshav.
Working paper at SSRN.
Accurate monitoring of natural resources at scale requires processing petabytes of satellite observations, yet persistent cloud cover, irregular observation patterns, and the scarcity of labeled training data limit conventional approaches.
We recently proposed the TESSERA self-supervised foundation model that generates 128-dimensional pixel embeddings from Sentinel-1 and Sentinel-2 time series at 10 m resolution with annual global coverage. TESSERA is trained on 800 million globally distributed pixels spanning 2017-2024 and uses a dual-branch Barlow Twin loss to learn spectrally and temporally decorrelated representations that preserve phenological signals and temporal dynamics despite being obscured by clouds, which eliminates the need for compositing, curve fitting or the extraction of hand-crafted feature sets.
Here, we evaluate TESSERA performance across five diverse downstream applications: crop type classification in Austria, wildfire burn area detection in California, canopy height estimation in Bornean tropical rainforests, above-ground biomass prediction in Finnish forests, and carbon market stocking indices in Brazilian agroforestry systems.
In all cases, TESSERA matches or exceeds the performance of task-specific supervised models. It achieves excellent label efficiency, typically reaching over 90% of final performance using less than 1% of available training data, making scalable monitoring for biodiversity conservation, climate policy, and ecosystem management possible even in resource-constrained environments.