The Rise of Earth Foundation Models

PRETRAINING

For years, Earth-observation models were built task by task. A flood classifier, crop classifier and burn-scar model could train on overlapping imagery while each learned its own internal picture of roads, water, vegetation, seasons and terrain from scratch. This was computationally expensive and conceptually repetitive.

Foundation models reverse the order. First learn a broad representation from very large quantities of Earth data, usually without requiring a label for every example. Then adapt that representation to a smaller downstream task. The expensive question becomes what structure is generally useful about Earth, not how to solve one taxonomy.

This is not a claim that one model understands every Earth process. It is a claim about reusable representation. If the same internal features help separate water from land, map burned areas, classify crops and detect change, there is little reason to relearn those features independently every time. The planet is complicated enough without charging compute for the same lesson repeatedly.

SELF-SUPERVISION

A model can learn from data before anybody supplies the answer

The central technical move is self-supervised pretraining. Instead of asking people to label millions of scenes, the training task is constructed from the data itself. A model may hide patches and learn to reconstruct them, predict one observation from another, distinguish matching from non-matching locations, or learn representations that remain stable across transformations.

Masked autoencoding is common in geospatial foundation models. The model sees only part of an image or time series and learns to reconstruct the missing information. To do this well, it has to learn recurring structures such as field geometry, water boundaries, urban texture, seasonal change and relationships between spectral bands. None of those needs to be named during pretraining.

The representation can later be adapted with a smaller labelled dataset. Fine-tuning updates some or all of the pretrained weights. Linear probing freezes the foundation and learns only a small output layer. Other workflows use the embeddings directly as features for a conventional model. The same foundation can therefore support users with very different amounts of compute and labelled data.

Generality is the test. A model that performs one benchmark beautifully but collapses when the task, sensor, geography or resolution changes is not much of a foundation. It is a specialist standing on a larger box.

NASA AND IBM

Prithvi trained on millions of Earth time series

NASA and IBM's Prithvi-EO-2.0 is a useful public example because both the weights and the training logic are inspectable. The 2024 technical report describes 300-million and 600-million-parameter models trained on about 4.2 million globally distributed time-series samples from NASA's Harmonized Landsat and Sentinel-2 archive at 30-meter resolution.

The model includes temporal and location embeddings, so the representation can use when and where an observation was made rather than treating every image patch as if it came from an anonymous planet. The training dataset was designed to broaden geographic and ecological coverage, with sampling across land-cover classes and hundreds of ecoregions.

In GEO-Bench evaluations reported by the team, the larger Prithvi-EO-2.0 model improved on its predecessor and compared favourably with several other geospatial foundation models across tasks and resolutions. More important than any single benchmark is the intended workflow: one pretrained representation can be adapted for land use, crop mapping, ecosystem monitoring, disaster response and other downstream applications.

In 2026, a version of Prithvi was demonstrated on two in-orbit computing platforms, which NASA described as the first deployment of a geospatial foundation model in orbit. That does not make the satellite generally intelligent. It does move a reusable model close enough to the sensor that downstream tasks no longer have to begin on the ground.

ALPHAEARTH FOUNDATIONS

Then the representation itself became a dataset

AlphaEarth Foundations takes a different approach. Rather than asking every user to run the foundation model over raw archives, Google DeepMind produced a global embedding field and released annual embeddings through Google Earth Engine. Each 10-by-10-meter terrestrial or coastal location is represented by a compact 64-dimensional vector.

Those vectors integrate information from multiple public Earth-observation sources, including optical imagery, radar, elevation-related data and climate information. Google describes more than 1.4 trillion embedding footprints per year in the released Satellite Embedding dataset. The representation has become something a user can query almost like any other geospatial layer.

That changes the economics of downstream mapping. A team with a relatively small labelled sample can train a lightweight model on top of the embeddings instead of downloading and processing a large multi-sensor archive. Google reported a 24% lower average error rate than the comparison systems in its evaluation and a much more compact representation.

The conceptual shift is larger than storage. Satellite imagery used to be the shared raw material and each application built its own representation. Now the representation itself can be shared infrastructure. The first expensive interpretation of the archive happens before the local question is even asked.

THE NEW INFRASTRUCTURE PROBLEM

A common foundation creates common dependencies

Reusing a foundation reduces the cost of building many applications. It also couples them. If a sensor condition, ecosystem, season or settlement pattern is poorly represented during pretraining, downstream systems can inherit the weakness together. Ten products built on one representation are not ten independent opinions.

Resolution creates another constraint. Many foundation models are trained with fixed patch sizes and internal scales. A representation useful for regional classification may be awkward for a task that needs metre-scale boundaries. ESA-supported work such as THOR is explicitly exploring compute-adaptive representations that can change internal resolution across Sentinel-1, Sentinel-2 and Sentinel-3 inputs rather than forcing every task through one fixed scale.

Evaluation is equally difficult. Foundation models are often compared across benchmark suites, but a benchmark averages over tasks chosen because labels already exist. Real deployments care about geographic transfer, rare events, new sensors, data scarcity and failure under changing conditions. Generality is easier to announce than to measure.

There is also an institutional question. Planet-scale pretraining requires data engineering and compute that relatively few organizations can afford. Open weights and public embedding datasets broaden access, but they also make many downstream systems dependent on representational choices made upstream. The foundation layer is beginning to resemble infrastructure, and infrastructure deserves scrutiny precisely because most users never rebuild it.

The foundation model does not remove specialization. It moves the shared part of specialization earlier, where everyone inherits it.