The Rise of Earth Foundation Models

PRETRAINING

For years, Earth-observation models were built one task at a time. A flood model, a crop model and a burn-scar model could train on largely the same imagery while each learned its own private picture of roads, water, vegetation, seasons and terrain from nothing.

This was expensive and, viewed from a slight distance, repetitive.

Foundation models reverse the order. First learn a broad representation of Earth from enormous quantities of data, mostly without labels. Then adapt it to a specific task with a much smaller labelled set. The expensive question becomes what is generally useful to know about the planet, rather than how to solve one particular list of categories.

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

SELF-SUPERVISION

A model can learn before anybody supplies the answer

The central trick is self-supervised pretraining. Instead of asking people to label millions of scenes, the training task is built out of the data itself. A model can hide parts of an image and learn to fill them in, predict one observation from another, tell matching locations from mismatched ones, or learn representations that stay stable when the input is altered.

Masked reconstruction is especially common for geospatial models. The model sees only part of an image or time series and must reconstruct the rest. Doing that well forces it to learn recurring structure: field geometry, water edges, urban texture, seasonal change and the relationships between spectral bands. None of it has to be named while it is being learned.

The representation is then adapted with a modest labelled dataset. Fine-tuning updates some or all of the pretrained weights. Linear probing freezes the foundation and trains only a thin output layer. Other workflows simply use the model's outputs as features in a conventional classifier. One foundation can therefore serve users with very different budgets of computing power and labels.

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

NASA AND IBM

Prithvi learned from millions of time series

NASA and IBM's Prithvi-EO-2.0 is a useful public example because both its weights and its training recipe can be inspected. The 2024 technical report describes models of 300 million and 600 million parameters trained on about 4.2 million globally distributed time-series samples from NASA's Harmonized Landsat and Sentinel-2 archive, at 30 metres.

The model includes embeddings for time and location, so it can use when and where an observation was made rather than treating every patch as if it came from an anonymous planet. Its training set was sampled across land-cover classes and hundreds of ecoregions to broaden what it saw.

In GEO-Bench evaluations reported by the team, the larger model improved on its predecessor and compared well with several other geospatial foundation models across tasks and resolutions. More important than any single score is the intended workflow: one pretrained representation adapted for land use, crop mapping, ecosystem monitoring, disaster response and whatever comes next.

In 2026 a version of Prithvi ran on two computing platforms in orbit, which NASA described as the first geospatial foundation model deployed in space. That does not make a satellite generally intelligent. It does put a reusable model close enough to the camera that a task no longer has to begin by waiting for the data to come down.

ALPHAEARTH FOUNDATIONS

Then the representation itself became a dataset

AlphaEarth Foundations takes a different route. Rather than asking everyone to run a foundation model over raw archives, Google DeepMind computed a global embedding field and released annual embeddings through Google Earth Engine. Every 10-by-10-metre patch of land and coast is summarised as a compact list of 64 numbers.

Those numbers combine information from several public 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, which can be queried much like any other map layer.

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

The bigger shift is conceptual. Satellite imagery used to be the shared raw material, and each application built its own interpretation of it. Now the interpretation itself is shared infrastructure. The first expensive reading of the archive happens before the local question has even been asked.

THE NEW INFRASTRUCTURE PROBLEM

A common foundation creates common dependencies

Reusing a foundation cuts the cost of building many applications. It also ties them together. If a sensor condition, ecosystem, season or settlement pattern was poorly represented during pretraining, every system built on top can inherit the weakness at once. Ten products built on one representation are not ten independent opinions.

Resolution is another constraint. Many foundation models use fixed patch sizes and internal scales, so a representation good for regional classification can be clumsy for a task that needs boundaries a metre wide. ESA-supported work such as THOR is exploring representations that adapt their internal resolution across Sentinel-1, Sentinel-2 and Sentinel-3 inputs rather than forcing every task through one scale.

Evaluation is just as hard. Foundation models are usually compared on benchmark suites, but a benchmark averages over tasks chosen because labels already existed. Real deployments care about transfer to new regions, rare events, new sensors, scarce data and failure when conditions change. Generality is easier to announce than to measure.

There is an institutional question too. Pretraining at planetary scale needs data engineering and computing that few organisations can afford. Open weights and public embeddings widen access, and they also make many downstream systems depend on choices made upstream by people they will never meet. The foundation layer is becoming infrastructure, and infrastructure deserves scrutiny precisely because most of its users never rebuild it.

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