Satellites See Every Pixel, But Not the Place: Why Earth Models Need a Geographic Grammar

A satellite can now capture every square meter of a country's land surface, yet still tell you almost nothing about the place it has photographed. This is the uncomfortable reality at the heart of a new critical review by Rajiv Ranjan and Shashank Tamaskar of Plaksha University (Ranjan & Tamaskar, 2026). The paper's argument comes down to a single, stinging distinction: a model can be geospatial — covering the whole Earth in pixels and coordinates — without being remotely geographic.

The authors draw a line between two capabilities the field has quietly collapsed together. Geographic completeness asks whether a model represents more than land surfaces and weather fields — whether it reaches the ocean's depths, the movement of species, the flow of trade, the institutions and inequalities that shape how people actually live. Geographic intelligence asks whether a model's outputs preserve place, scale, relations, process, uncertainty, and the limits of what can responsibly be claimed. The first concerns what exists inside a model; the second concerns what may be said about it. Confuse the two, Ranjan and Tamaskar argue, and you get systems that are impressively broad and stubbornly shallow.

The Science

This is not an experiment with a dataset and a benchmark. It is a critical integrative review — a synthesis of a sprawling, fragmented literature into a new conceptual frame. The authors combed through a decade of research spanning arXiv preprints, peer-reviewed articles, and institutional reports, covering Earth foundation models, ocean forecasts, multimodal encoders, geospatial reasoning agents, ecology, economics, and the long traditions of GIScience (geographic information science) and human geography. Because the field's model comparisons are notoriously fickle — varying in sensors, training corpora, resolution, and evaluation protocols — Ranjan and Tamaskar deliberately avoided any single numerical ranking. Instead, they ask what each family of models makes possible, and which requirements remain untested.

Their central move is literary in the best sense: they borrow the idea of a grammar and make it computational. Just as a sentence needs more than a vocabulary of words, a geographic model needs more than a collection of georeferenced layers. The proposed grammar has three parts. Geographic primitives — the places, entities, events, attributes, relations, and flows that populate a model's world. Dynamic operators — the processes, transitions, trajectories, drivers, and interventions that change things over time. And contextual constraints — scale, spatial support, uncertainty, provenance, and validity, which govern what can be legitimately claimed.

The whole grammar is captured in a typed formal system,

$$\Gamma =(\mathcal{V},\mathcal{R},\mathcal{P},\mathcal{S},\mathcal{T},\mathcal{A},\mathcal{U}),$$

where $\mathcal{V}$ holds places and events, $\mathcal{R}$ holds relations like adjacency and flow, $\mathcal{P}$ holds transitions, $\mathcal{S}$ and $\mathcal{T}$ set spatial and temporal support, $\mathcal{A}$ represents natural and human agency, and $\mathcal{U}$ records uncertainty and provenance. A model counts as geographic to the degree its inferences respect the constraints among these types — a flow connects admissible entities, a process operates at a defensible scale, an explanation states its evidence and its limits.

The grammar's power is that it supports modularity. An ocean model can keep its depth coordinates, an atmosphere model its pressure levels, an ecological model its species networks, an economy its trade graphs. "Shared identifiers and typed relations let these systems refer to the same changing places without forcing every domain onto one grid." Grammar, in other words, is an interoperability layer, a reasoning structure, and an epistemic constraint all at once.

What They Found

The heart of the review is a table of seven "structural gaps" — seven ways the field's convenient computational shortcuts quietly substitute for the richer geographic concepts they approximate. Each is useful; none is equivalent. Ranjan and Tamaskar state them as blunt propositions.

Location is not place. Coordinates tell you where something was recorded. Place — constituted by environmental character, history, function, meaning, and connection — tells you why it matters. A model built on location embeddings can perfectly separate two visually similar landscapes yet have no idea why one floods, one thrives, and one is abandoned.

Resolution is not scale. Resampling the same image to different pixel sizes does not establish scale-aware reasoning. Pattern and explanation can flip with scale — a relationship that holds at field level may reverse or disappear at regional level. Yet most models treat a pixel grid as if it captured every relevant scale at once.

Temporal encoding is not process understanding. Timestamps and attention mechanisms help models exploit seasonality. They do not identify drivers, separate disturbance from recurrence, or predict how a trajectory changes when forcing changes. Reconstructing the past under a known climate is weak evidence of understanding the future under a novel one.

Multimodality is not geographic synthesis. A model can fuse optical, radar, climate, and socioeconomic data while learning only statistical co-occurrence — it never distinguishes driver from state, observation from intervention, exposure from outcome.

Prediction is not explanation. High accuracy establishes association under a particular evaluation design. It does not establish mechanism or cause. A model that forecasts crop stress from imagery and weather may still fail to tell you whether the culprit is water deficit, disease, management, or a sensor artefact. "Fluent language," the authors warn, "cannot upgrade association into explanation."

Global coverage is not geographic universality. A globally distributed training corpus can still privilege data-rich regions, common land classes, particular sensors, and institutional definitions. The review cites two audits — PANGAEA and EarthShift — showing that foundation models don't consistently beat supervised baselines, that nominally similar benchmarks disagree substantially, and that performance degrades under realistic geographic, temporal, scale, and sensor shifts.

Earth-surface coverage is not geographic completeness. The most damning gap. A model can cover every land pixel and still represent only a selected exterior of the planet. The three-dimensional ocean, the organization and movement of life, economic networks and flows, institutions, meanings, and inequalities — "the imbalance," the authors note dryly, "favours terrestrial surfaces and gridded physical fields."

Seven propositions follow from these gaps, each phrased as a testable requirement. Transfer between visually similar but functionally different places. Aggregation consistency across scales and boundaries. Multiple forecast horizons under unseen forcing combinations. Missing-modality robustness. Explanations tied to locatable evidence with claims labeled as descriptive, predictive, attributive, causal, or counterfactual. Region-, time-, sensor-, and scale-held-out tests with disaggregated performance for rare classes and data-poor regions. And cross-domain audits validating the learned couplings among land, atmosphere, ocean, biosphere, economy, and society.

Why This Changes Things

Read one way, this review is a diagnosis of a field's blind spots. Read another, it is a quiet reordering of what success even means.

The authors are careful not to dismiss the extraordinary progress of recent years. Earth-observation foundation models like AlphaEarth and Prithvi turn vast unlabeled satellite archives into reusable representations. TESSERA learns pixel-wise embeddings from irregular Sentinel time series, notable for open weights and lightweight, accessible annual 10-meter representations. Aurora spans weather and broader environmental prediction. WenHai represents depth-aware upper-ocean dynamics and air-sea forcing. TerraMind and SkySense++ push generative multimodality across Earth observations. Reasoning agents like Earth AI, Geo-CoT, and TerraLogic coordinate tools, ground analysis in perceptual evidence, and decompose questions into hierarchical workflows.

But the review's contribution is to insist these are complementary specialists, not rivals for some imagined universal model. The question the field should be asking is not "which one is best" but "what can each contribute to a shared geographic state." An ocean model should not be expected to explain migration; a weather model should not be expected to model culture. Specialist systems remain essential. The error arises only when a partial observational scope is mistaken for a complete Earth ontology, or when benchmark performance is treated as evidence of place-based understanding.

The review's most consequential claim concerns the gap between technical progress and real consequence. "A reusable embedding or benchmark gain becomes consequential only after it survives geographic shift, represents a decision-relevant variable, reaches a user at the required time, communicates uncertainty, fits an institutional workflow, and improves a measured outcome." Evidence that this chain actually works remains strikingly thin. This is not, the authors insist, evidence that GeoAI lacks value. It shows that most evaluations stop at representation or prediction, while real deployment is an out-of-distribution and socio-technical problem — a map must align with thresholds, calendars, costs, ownership, correction, and accountability.

Consider what this means for a farmer, a coastal planner, or a health official who receives a model's output. A perfectly accurate prediction is useless if it arrives a day late, at the wrong resolution, without uncertainty, or unaligned with the institutional calendar on which decisions actually run. The review reframes success: "A geographic model should be judged not only by what it maps, but by whether its use produces a defensible improvement for particular places, populations, and ecosystems."

What's Next

The paper's destination is a concept it calls the Living Geography Model (LGM) — the proposed mature form of geographic intelligence.

An Intelligent Earth Model represents, predicts, or reasons about Earth phenomena. A Geographic model organizes those capabilities through place, scale, relation, process, and human-environment interaction. An LGM is the fuller ambition: a federated, continually updated, multiscale system of coupled places, connecting land, atmosphere, ocean, biosphere, economy, and society.

The authors are explicit that this is not a single universal network swallowing everything. "The objective is not one universal network," they write — federation is the point. Specialist models would remain specialists, but they would share place identities, spatial and temporal support, relations, provenance, and uncertainty — the connective logic that geography supplies. Rather than collapsing every domain onto one grid, such an architecture lets each system speak its own geometry while agreeing on what places they refer to, at what scales those claims hold, and with what uncertainties.

The honesty of the review is refreshing. Seven propositions are designed, in the authors' own words, to be "citable, contestable, and testable" — a deliberate attempt to prevent "a recurring error in which a convenient computational proxy is treated as equivalent to the geographic concept it only partially measures." And the paper concedes its own boundaries: as a purposive rather than exhaustive review, it does not justify claims about the frequency of every feature across the complete literature. The gaps it names are offered as hypotheses to be tested, not verdicts to be accepted.

What this opens up is a research agenda that is simultaneously more humble and more ambitious than current practice. More humble, because it asks models to state what they cannot do — which domains are unrepresented, which scales are unsupported, which claims are beyond their grounds. More ambitious, because it asks for models that don't just map where things happen but explain how places are connected, why they change, what may happen next, and how interventions might alter their futures.

The deepest lesson may be this: as our models grow larger and cover more of the planet, the binding constraint is no longer data or compute. It is whether we have the grammar — the conceptual and computational logic — to know what a model is actually entitled to claim about the places it portrays. Geography, Ranjan and Tamaskar argue, is that grammar. The payoffs for getting it right are not academic. They are the difference between a prediction that sits in a paper and a decision that changes lives in a particular place, at a particular scale, with a defensible measure of what it will do.