Space is continuous, but computation needs addressable units. Spatial tokenisation is the process of turning locations, coordinates, cells, regions, routes and places into discrete or model-ready representations that a system can compare, retrieve, predict and reason over.
A latitude–longitude pair, a map tile, a postal code, a hexagonal cell, a neighbourhood name and a building ID can all point to space at different granularities. None is the place itself. Each is a representation chosen for a job.
This article extends the eduKateSingapore Representation and Tokenisation series into spatial reasoning.
The Spatial Representation Route
WORLD LOCATION → COORDINATE / PLACE / SENSOR OBSERVATION → REFERENCE SYSTEM → GRID / REGION / PLACE ID → SPATIAL TOKEN OR EMBEDDING → NEIGHBOURHOOD / DISTANCE / TOPOLOGY → MODEL → ROUTE / SEARCH / PREDICTION → MAP / ACTION → WORLD RETURN
1. Coordinates Are Representations, Not Places
A coordinate describes a position inside a reference system. Without the reference system, the numbers are incomplete.
The same physical place can have different coordinate representations under different datums and projections.
2. Reference Systems Define the Spatial Contract
Latitude and longitude depend on an Earth model. Projected coordinates depend on a projection. Local engineering coordinates can depend on a site-specific origin.
Spatial tokens are trustworthy only when the reference system travels with them.
3. Precision Is Not Accuracy
A coordinate can contain many decimal places and still be wrong because the sensor, geocoder or datum transformation is inaccurate.
More digits preserve representational precision, not necessarily truth.
4. Grids Turn Continuous Space Into Cells
A grid divides a spatial domain into discrete cells. Every point inside one cell can share one cell identifier.
The cell ID becomes a spatial token.
5. Cell Size Controls Spatial Granularity
Large cells compress many locations into one token. Small cells preserve local detail while creating a much larger vocabulary of possible places.
Resolution is therefore a vocabulary-versus-detail trade-off.
6. Grid Boundaries Are Artificial
Two homes metres apart can fall into different cells while two places far apart inside a large cell share one token.
The boundary is computational, not geographical truth.
7. Hierarchical Grids Provide Multiple Resolutions
A coarse parent cell can be divided into finer child cells. The same location can therefore have tokens at neighbourhood, street and building-like scales.
Hierarchical spatial tokenisation preserves a route between broad context and fine detail.
8. DGGS Makes Hierarchical Spatial Addressing Explicit
The Open Geospatial Consortium describes Discrete Global Grid Systems as reference systems that organise the globe into hierarchies of addressable zones. OGC’s current DGGS standards formalise structured geometry, hierarchical refinement and unique zonal identifiers.
The important representation lesson is broader than any one grid: a region of space can receive a stable discrete address.
9. Equal-Area Cells Protect Some Analyses
If spatial cells vary greatly in area, raw counts per cell can be misleading. Equal-area schemes make each cell represent comparable surface area.
The geometry of the token affects statistics computed over it.
10. Hexagons, Squares and Triangles Have Different Neighbourhoods
Cell shape changes adjacency, distance approximation and directional bias.
Spatial token geometry is therefore part of model design, not merely visual styling.
11. Geohash-Style Encodings Turn Coordinates Into Strings
Coordinate regions can be encoded into compact strings whose prefixes correspond roughly to broader areas.
Such strings make hierarchical geography usable in databases and sequence models, while inheriting boundary and shape limitations from the encoding.
12. Prefix Similarity Is Not Always Physical Proximity
Two nearby points across an encoding boundary can have dissimilar prefixes. Another pair sharing a long prefix can be farther apart inside the same cell.
String similarity should never be mistaken for exact spatial distance.
13. Places Are Semantic Regions
“Singapore”, “Punggol”, “Changi Airport” and “a classroom” are not just coordinates. They are named places with social, administrative or functional meaning.
Place tokenisation adds entity identity above geometry.
14. Place Names Need Entity Resolution
Several cities can share one name. A business can have many branches. A road name can occur in multiple countries.
Place identity therefore needs disambiguation, jurisdiction and often a canonical ID.
15. Administrative Boundaries Change Over Time
Districts, electoral boundaries and planning areas can be redrawn. A spatial token that meant one region in 2020 may represent another in 2030.
Temporal versioning belongs with administrative geography.
16. Natural Regions Are Fuzzier Than Administrative Regions
A coastline, wetland, commercial district or cultural neighbourhood may not have one universally agreed boundary.
Spatial representations should preserve uncertainty where boundaries are interpretive.
17. Points, Lines and Polygons Encode Different Spatial Objects
A bus stop can be represented as a point, a road as a line and a park as a polygon.
The geometry type determines which spatial operations make sense.
18. Centroids Are Convenient and Potentially Misleading
Replacing a region with one central point makes storage and comparison easier. It also discards shape, extent and internal variation.
A centroid is a compressed representation, not the region itself.
19. Bounding Boxes Are Coarse Spatial Tokens
A rectangle can bound an object or region compactly. It preserves rough extent but includes space not belonging to the object.
Bounding boxes trade geometric fidelity for simplicity.
20. Topology Can Matter More Than Coordinates
For some tasks the important question is whether regions touch, contain, overlap or connect—not their exact coordinates.
Topological relations are higher-level spatial tokens.
21. Adjacency Creates a Spatial Graph
Cells, road segments or stations can become nodes connected by neighbour or route edges.
This connects spatial tokenisation to Graph Tokenisation.
22. Distance Is Representation-Dependent
Straight-line distance, road distance, walking time and transit time can disagree dramatically.
The right distance metric depends on the receiver’s job.
23. Route Distance Is a Network Property
Two points separated by a river may be physically close and operationally far apart.
Navigation requires topology and transport constraints, not coordinates alone.
24. Trajectories Are Sequences of Spatial Tokens
A moving object can be represented by timestamped coordinates, visited grid cells or named places.
The trajectory becomes a spatial–temporal token sequence.
25. Sampling Rate Changes Trajectory Fidelity
Recording a position every second preserves detail. Recording once every hour can miss detours and stops.
Temporal sampling and spatial granularity interact.
26. Map Matching Converts Noisy Coordinates Into Network Tokens
GPS observations can be mapped to likely road segments or rail lines.
The system replaces noisy point observations with a higher-level route representation.
27. Map Matching Is Inference, Not Measurement
Near parallel roads or weak signals can make several routes plausible.
The matched segment should preserve confidence and the original sensor evidence.
28. Spatial Embeddings Learn Neighbourhood Similarity
Grid cells, places or routes can be mapped into vectors based on co-visitation, nearby context or graph structure.
The embedding captures learned spatial similarity while the canonical spatial ID preserves exact identity.
29. Proximity and Similarity Are Different
Two airports on opposite sides of the world may have similar embeddings because they serve similar functions. Two adjacent land parcels may have very different functions.
Spatial closeness is only one relation.
30. Raster Maps Are Spatially Tokenised Images
A raster divides space into a regular grid of pixels or cells, each storing a value such as elevation or land cover.
Raster resolution determines the smallest spatial pattern the dataset can express.
31. Vector Maps Preserve Geometry Differently
Vector data stores points, lines and polygons rather than one value per cell.
Raster and vector formats are alternative spatial representations, not better and worse versions of one another.
32. Spatial Aggregation Creates the Modifiable Areal Unit Problem
Statistical patterns can change when the same observations are grouped into different regions or scales.
Spatial token boundaries can therefore change apparent correlations.
33. Privacy Can Require Coarser Spatial Tokens
Exact home coordinates can expose individuals. Aggregating to broader cells or regions can reduce privacy risk.
The coarser representation sacrifices precision to protect people.
34. Coarsening Does Not Guarantee Anonymity
Rare trajectories or combinations of locations can still identify individuals even when individual points are blurred.
Spatial privacy must be evaluated against linkage risk, not cell size alone.
35. Geocoding Is a Representation Bridge
Geocoding converts an address or place string into coordinates or a spatial entity. Reverse geocoding performs the opposite transformation.
Both are inference processes that need confidence and source provenance.
36. Address Parsing Is Not Location Truth
A perfectly parsed address can still refer to a demolished building, a moved entrance or an ambiguous unit.
World return requires current location evidence.
37. Spatial Search Uses Several Token Scales
Search can use place names, administrative regions, coordinate radii, grid cells and route networks at once.
Multiple spatial representations improve retrieval because they fail differently.
38. Spatial Language Is Relative
“Near”, “north of”, “inside” and “across from” depend on scale, reference frame and context.
Natural-language spatial reasoning must connect symbolic relations back to geometry.
39. Spatial Models Need Freshness
Roads close, buildings open, boundaries move and transport networks change.
A spatial token can remain syntactically valid while its world referent becomes stale.
40. Versioned Spatial Data Preserves Historical Truth
A route valid in 2024 may not be valid in 2026. Historical analysis should use the map edition appropriate to the event time.
Spatial representation needs time-aware provenance.
41. The Spatial Tokenisation Audit
- What real-world spatial object is being represented?
- What coordinate reference system applies?
- What spatial precision and accuracy are required?
- Is the unit a point, cell, line, polygon or named place?
- What cell size or hierarchy level is used?
- How do grid boundaries affect neighbours?
- What distance metric matches the task?
- Are place names linked to canonical spatial entities?
- Do administrative boundaries need a date/version?
- How are uncertainty and geocoding confidence preserved?
- Are trajectories sampled densely enough?
- Does aggregation distort the analysis?
- What privacy coarsening is required?
- Can every derived spatial token return to source geometry?
- Is the spatial data fresh enough for the decision?
42. What Students Should Remember
- Coordinates only make sense inside a reference system.
- Grids turn continuous space into discrete cells.
- Cell size controls spatial granularity.
- Place names add semantic identity above geometry.
- Distance, topology and route connectivity are different spatial relations.
- Trajectories combine space and time.
- Spatial boundaries can change apparent statistics.
- Spatial tokens need provenance, freshness and world return.
43. The Deep Principle
Spatial tokenisation gives the machine handles on continuous geography. Every handle chooses a scale and therefore chooses what local differences remain visible.
A place becomes computationally useful when it can be addressed, related and returned to the world—but the address is never the place, and the grid is never the geography.
Continue the Representation & Tokenisation Series
- Document Layout Tokenisation | How Pages, Headings, Tables, Boxes and Reading Order Become Structure
- Biological Sequence Tokenisation | How DNA, RNA and Proteins Become Model-Ready Units
- Workflow Tokenisation | How Processes Become Steps, Actions, States and Decisions
- Canonical owner: World Representation & Cognitive Tools