Geospatial foundation models (GFMs) are getting bigger. More sensors, more modalities, more parameters. The assumption seems obvious: if a model is meant to be foundational, it should know as much about the Earth as possible.
At Spheer, we are taking a different approach.
Spheer FM is deliberately constrained. It is focused on nature and biodiversity, trained for Europe, and designed around the temporal behaviour of the Earth's surface. It gives up some universality to learn representations that are more useful for the ecological systems we care about.
That raises a broader question: can a foundation model become more valuable by becoming less universal? (spoiler: yes)
Foundational does not mean general
The implicit promise in ever-larger GFMs is that generality wins. But a foundation model does not have to be general-purpose. It can be foundational within a domain: pretrained once on large amounts of unlabelled data, then reused without finetuning across many downstream tasks in that domain.
Generality has a cost. A general-purpose GFM has to represent deserts, rainforests, European farmland and tropical coastlines; urban infrastructure and natural ecosystems; different climates, seasons, sensors and acquisition conditions. If your downstream task is identifying tree species, classifying crops or characterising habitats, most of that variation is irrelevant — and a model built to represent everything everywhere spends capacity on differences that do not matter for the task at hand.
Smaller, focused models are also easier to train, run and iterate on: Spheer FM has less than 1% of AlphaEarth's parameters. That matters beyond cost. When a model is cheap to train, its design choices can be made to correspond directly to the physical processes we care about and we can afford to build several specialised models instead of one that has to serve everything.
A pixel is a signal, not a part of a picture
Many GFMs are, at their core, computer vision models, and they inherit computer vision's assumptions: images are divided into patches, nearby pixels provide context, spatial relationships explain what is being observed.
For many Earth Observation problems, that makes sense. For ecological ones, there is another way to look at the data. A patch of grassland is interesting not because of what its neighbours look like, but because of what happens to its reflectance across the year: vegetation emerges, grows, flowers, gets harvested or mown, and disappears. A pixel is a signal through time.
That is why Spheer FM produces yearly embeddings. A calendar year maps onto a complete growing season, with winter sitting naturally at the boundary between one annual embedding and the next.
And it is why we deliberately limit spatial context instead of maximising it. By reducing the opportunity to rely on nearby pixels, field shapes or surrounding infrastructure, we push the model toward the temporal signal. It cannot recognise the shape of a field. It cannot use a neighbouring road as a clue. And it becomes harder for it to memorise where particular patterns occur on the Earth's surface.
This is a radical simplification, and that is exactly the point. The restriction is not a limitation we reluctantly accept. It is the inductive bias we want.
Does specialisation actually work?
We have been benchmarking Spheer FM against other prominent GFMs: Google’s AlphaEarth and Cambridge's TESSERA, on two European benchmarks, using the precomputed embeddings as-is, without retraining any backbone.
- The first is TreeSAT-AI, a German tree-species classification benchmark.
- The second is PASTIS, a French crop-classification benchmark.
Spheer FM leads on both benchmarks, and by a wide margin on the harder TreeSAT levels. The crop result is especially interesting, because crop classification was not the primary target of our pretraining design. And the fact that a simple KNN performs this well on our embeddings suggests the decision boundaries are simpler and the embedding space more logically structured for in-domain tasks.
The takeaway
Spheer FM is not the model you would choose to analyse urban growth in Southeast Asia or geological features in the Sahara. It is not trying to be. But on the problems it was built for, it outperforms models that have seen far more of the planet.
GFMs should not be judged by how big they are, or how much of the Earth they saw during training, but by how useful their representations are for the problems we actually want to solve.
Perhaps the best foundation model for a problem is not the one that knows the most, but the one that knows what matters.
