INFORMATION SCALECHANGING SCALE…VIEW / READYING
GEOAI / LIMITS OF GEOGRAPHIC REASONING

THE LIMITS OF AI / GEOAI

The Limits of AI:Geospatial Foundation Models
Do Not Understand Geography

Geospatial foundation models can identify surface objects and spatial patterns.
But spatial representation is not the same as geographic understanding.

Here, “do not understand” means that current models still lack robust, scale-aware and mechanism-oriented geographic reasoning.

00 / 05
CONTINUE

01 / OBSERVATION

Models first learn observational representations.
Not geographic mechanisms themselves.

Imagery, maps and trajectories are observations of the world.
They contain geographic information, but are conditioned by sensors, scale and sampling.

Recognizing spatial patterns
is not the same as explaining the processes that produced them.

MOVE / Pixels provide observational evidence; process, relation and history still require inference.

EARTH OBSERVATION / SENSOR × SCALE × TIME
PIXELWORLD?
01 / 05
CONTINUE

02 / SPATIAL GENERALIZATION

The real test is whether a model
generalizes across regions.

Random splits often leave training and test samples in similar spatial environments and can overstate performance.
Spatial holdout is closer to deployment: the model must face unseen regions and distribution shift.

The key geographic capability
is robust inference across places.

EarthShift reports roughly 15–20% average degradation under real spatial distribution shifts. See Sources.

TRAIN
TEST
SCHEMATIC SCORE
92
Conceptual interaction, not the result of a single model; it illustrates the evaluation gap between random split and spatial holdout.
02 / 05
CONTINUE

03 / PLACE & SCALE

Coordinates locate an object.
They do not define a place.

Place is constituted by location, scale, spatial relations, environmental processes and history. Geographic claims are therefore scale- and context-dependent.

Change the scale,
and relations, dominant processes and valid conclusions can change.

DRAG SCALE / The same spatial object participates in different relation structures at different scales.

PLACE = LOCATION × SCALE × RELATION × PROCESS × HISTORY
SCALE1 KM
03 / 05
CONTINUE

04 / SPACE ≠ GEOGRAPHY

Spatial representation
is not geographic understanding.

3D reconstruction and viewpoint transformation describe geometric consistency.
Geographic explanation also requires regional differentiation, process mechanisms and spatial dependence.

A spatial model answers where things are and what geometry they have.
Geographic intelligence must also explain why they occur here and why places differ.

MOVE / Viewpoint changes geometric representation; it does not automatically explain the mechanisms of place.

PLACESCALERELATIONPROCESSHISTORY
3D SPATIAL WORLDGeometry can be reconstructed.
Geographic relations still require reasoning.
04 / 05
CONTINUE

05 / GEOGRAPHIC REASONING

The central limitation of current geospatial foundation models
is geographic reasoning.

Reliable geographic intelligence must state its scale and domain of validity, recognize spatial dependence and regional heterogeneity, and determine when conclusions can transfer and when they cannot.

01 / SCALEState scale and domain of validity
02 / RELATIONRecognize dependence and heterogeneity
03 / TRANSFERTest geographic transfer

Core claim: spatial representations can be learned; geographic understanding must be tested across scales, regions and mechanisms.

05 / 05
END / SOURCES