Foundation geo models

Large language models changed how machines handle human knowledge. Not by memorizing the internet, but by learning compressed representations of structure, relationships, behavior, and meaning from massive amounts of data. That was the real breakthrough. The model stopped treating information as isolated documents. It began modeling the latent structure underneath them.

the leap was not memorization. it was representation.

Earth observation is approaching a similar transition. Today most EO systems behave like retrieval systems: search imagery, run detections, compare scenes, generate analytics. But planetary-scale sensing infrastructure now produces enough continuous observational data for a different kind of system to emerge: foundation geo models, trained not merely on images but on the evolving physical state of Earth itself. The early ones are already public. IBM and NASA released Prithvi on Hugging Face in 2023, trained on harmonised Landsat and Sentinel-2 imagery, and anyone can download the weights. They are early and narrow, but they are the shape of the thing. As they mature, Earth observation stops being about remote sensing and starts being about machine understanding of physical reality.

Early computer vision focused on recognition: detect a ship, segment a road, classify a building. Most EO AI still works this way. It is useful and it is limited, because reality is dynamic, relational, and temporal rather than a pile of disconnected objects.

Ports influence supply chains. Weather influences agriculture. Infrastructure influences economics. Energy systems influence geopolitics. The intelligence is rarely inside a single observation. It lives in relationships across time. Foundation geo models try to learn those relationships directly, moving past “what is visible” to “what is happening”, “what caused it”, and “what happens next”. That is the step from perception to reasoning.

a port alone is a noun. a port through time is a system.

The physical world can’t be understood from a single sensing modality. Optical imagery alone is insufficient. SAR alone is insufficient. Thermal sensing alone is insufficient. Reality expresses itself across multiple physical dimensions simultaneously.

A future geo model ingests optical imagery, hyperspectral observations, SAR, thermal signals, elevation data, weather, maritime telemetry, economic indicators, and infrastructure graphs. Traditional GIS kept every one of those in its own fragmented system.

Foundation geo models unify them into shared representations of planetary state. The model stops thinking in raw pixels and starts learning representations of infrastructure, environmental dynamics, industrial activity, and physical causality. The world becomes machine-representable as a continuously evolving latent structure.

Most current EO systems are surprisingly forgetful. They analyze imagery episodically rather than continuously. But intelligence requires memory. A human analyst is valuable for historical context, not for inspecting a single image. They know what existed before, what changed slowly, which patterns repeat, and which anomalies matter.

Future geo models need persistent temporal memory, and archival storage isn’t the same thing. A planetary intelligence system holds evolving representations of infrastructure development, agricultural cycles, hydrological behavior, industrial expansion, and urban growth, and updates them as the world moves.

That produces machine memory of planetary history at civilization scale. The model learns how the planet behaves through time, which is something no archive has ever done.

archives remember what happened. memory remembers what it means. the difference is everything.

Perception alone isn’t intelligence. Reasoning is. A foundation geo model becomes valuable when it starts connecting observations into operational understanding. Instead of “detect flooding”, the system reasons: “flooding risk is increasing because rainfall intensity, upstream reservoir saturation, historical drainage behavior, and infrastructure vulnerability patterns are converging.” That isn’t image analysis. It is planetary reasoning.

Shipping anomalies point to supply chain disruption. Reservoir decline points to agricultural instability. Construction acceleration signals industrial growth, thermal signatures signal operational activity, and environmental stress signals migration pressure. The model starts behaving like a simulation engine for civilization, probabilistically aware rather than perfectly predictive.

The model sees the planet and expects things of it. It develops expectations about how the planet behaves, and the intelligence comes from the expectations it gets wrong.

the model isn’t watching the world. it’s running a prior against it. the anomaly is where the prior breaks.

Once systems maintain temporal understanding at sufficient scale, prediction follows. Not because the system knows the future, but because physical systems contain patterns. Climate systems evolve dynamically but not randomly. Infrastructure development follows economic signals. Agricultural stress follows environmental conditions. Human logistics exhibit behavioral regularities.

A capable geo foundation model could support climate trajectory forecasting, infrastructure growth prediction, supply chain disruption modeling, environmental risk estimation, and urban expansion simulation. That converges with digital twin work, except the twin isn’t one city. It is a continuously updated probabilistic model of Earth.

Earth as a digital twin you don’t have to build. it builds itself, frame by frame.

This changes the role of Earth observation entirely. EO stops functioning purely as a measurement layer. It becomes part of a planetary prediction architecture.

Climate change may become one of the largest drivers of planetary intelligence infrastructure, because climate is a systems problem. No single dataset explains it fully. Temperature alone is insufficient. Imagery alone is insufficient. Weather alone is insufficient. Climate understanding requires integrating atmospheric systems, water cycles, agriculture, infrastructure, energy systems, population movement, ecological behavior, and economic adaptation.

Foundation geo models suit this because they work across modalities and timescales at once. The system can learn drought emergence, ecosystem degradation, urban heat stress, flood vulnerability, and agricultural decline from direct observation rather than simulation equations alone. That is a new category of climate infrastructure: observation-trained planetary intelligence.

Civilization expresses itself physically: roads, ports, factories, power grids, pipelines, warehouses, cities. Infrastructure is effectively the visible operating system of human civilization. Foundation geo models may become the dominant systems for understanding how that operating system evolves globally — not by manually analyzing maps, but by continuously learning planetary infrastructure behavior directly from observation streams.

the infrastructure is the API. roads, ports, grids. that is how civilization exposes itself to a sensor.

An advanced model could infer economic acceleration, industrial decline, energy transitions, supply chain fragility, and urbanization momentum from continuously updating planetary signals. That is why Earth observation converges with intelligence infrastructure rather than mapping software. The value sits in understanding civilization, not in the images.

Language models learned human knowledge from human-generated text. Foundation geo models may learn physical-world intelligence from planetary observation. That is a different category of machine understanding: reality rather than language.

Language is an abstraction humans built about the world. Planetary sensing observes the world directly. So models can train on physical state transitions rather than human descriptions of them. The implications run past Earth observation into robotics, autonomous systems, climate adaptation, infrastructure planning, and disaster response. Any system that depends on understanding physical reality ends up depending on geo foundation models underneath.

The future EO stack may not be defined by satellites, maps, imagery, or even sensing. It may be defined by foundation models that hold an evolving machine understanding of the physical world: models that remember planetary history, reason about change, simulate trajectories, and read infrastructure, climate, and economics together.

Remote sensing was the name for an industry that fetched pictures. What comes next answers questions about the planet, and it will need a different name.