Field Work
The data doesn't
collect itself.
Every benchmark starts with someone going there.
Climate AI is only as good as the data it trains on. But for most of the planet, especially in the Global South, that data is missing, coarse, or unverifiable. Nobos builds benchmark datasets for geospatial AI: ground-truth measurements, structured for machine learning, with full provenance.
Elevation, soil, biomass, land cover — the environmental datasets powering climate models today were built from what was easy to reach.
Large parts of the planet have never been directly measured. Models trained on this data don't just have errors. They have blind spots baked in, with no way to know how large they are.
The data doesn't
collect itself.
Every benchmark starts with someone going there.
Field campaigns in hard-to-reach environments, turned into ML-ready benchmark datasets with STAC, GeoParquet, and GeoJSON.
The question we answer: how accurate is the data you're already using?
We identify where existing data fails your use case and define exactly what needs to be measured.
Deploy our global network of field scientists, explorers, and domain experts.
Design collection protocols for accuracy, reproducibility, and ML compatibility.
Collect in the field, including environments no satellite can reach.
Structure, validate, and deliver datasets ready for machine learning.
Differential GPS surveys accurate to ±6cm. The question: how wrong is the elevation data everyone is already using? The answer: a baseline model showed an average error of 368 meters against three major elevation datasets.
184 mountains measured globally, and counting. See the NY Times story here ↗
Want to take a look at the data? →Be the first to know when new datasets drop.