2.2 Scale, Resolution, and Generalisation

Key takeaways

• "Scale" means at least three different things in GIS: cartographic scale, ground resolution, and zoom level.
• Detail that looks great at one scale is noise at another; generalisation bridges the gap.
• Matching data resolution to question resolution is a precondition for valid analysis.

Introduction

"Scale" is the single most overloaded word in geography. It can mean a ratio on a paper map, the size of the smallest feature your satellite can resolve, or the slippy-map zoom level in a web app. Mixing these up causes real bugs — literally, in the sense that an analyst might use 10 m resolution imagery to verify a claim only supported by 250 m data. This lesson untangles the three meanings and introduces the process of generalisation that lets data designed for one scale be usable at another.

Three meanings of "scale"

1. Cartographic scale (representative fraction)

On a paper map, scale is a ratio: 1 cm on the map equals N cm on the ground. Common notations:

• Representative fraction: 1:50 000 — one unit on the map is 50 000 units in reality.
• Graphic scale bar: a printed bar with unit labels (e.g., "0—5—10 km").
• Verbal: "one inch to the mile".

A large-scale map shows a small area in detail (1:1 000). A small-scale map shows a large area with less detail (1:10 000 000). The numerical intuition is counter-intuitive: large scale = large fraction = more detail.

2. Ground resolution (spatial resolution)

For raster data, resolution is the linear size of one cell on the ground. Sentinel-2 multispectral bands have 10 m resolution (each cell represents a 10×10 m patch of ground). A Worldview-3 satellite image can have 31 cm resolution. Lower numerical resolution = finer detail.

For vector data, "resolution" is fuzzier. A common proxy is the smallest feature size captured — are all footpaths represented, or only highways? Another is the positional accuracy — ±1 m? ±50 m?

3. Zoom level

On slippy web maps, zoom level is an integer: 0 (whole world in one tile) to 22+ (a small garden in one tile). Each increment doubles linear resolution and quadruples the number of tiles. At zoom z, the world is a 2^z × 2^z grid. Zoom level is an approximate proxy for cartographic scale — the relationship depends on screen DPI.

Matching data to question

A simple rule: the resolution of the data must be finer than the question you're asking.

• Counting parking spaces → need sub-metre imagery.
• Mapping deforestation of rainforest patches → 10–30 m resolution is enough.
• Continental climate modelling → 1–10 km is fine.
• Asking about a single building → you cannot answer that with a 250 m raster.

When you inherit a dataset, the first metadata to check is its resolution. If it's coarser than your question, stop before you generate misleading answers.

Generalisation — the art of simplifying

When you zoom out, you cannot show every feature. Cartographic generalisation is the process of simplifying data for display at coarser scales. Classic operations:

• Selection — omit minor roads at small scales.
• Simplification — reduce the number of vertices in a coastline using Douglas–Peucker or Visvalingam algorithms.
• Smoothing — round off jagged boundaries.
• Aggregation — merge adjacent polygons (e.g., combine all residential zones into a single "urban" polygon).
• Displacement — move features apart when they'd overlap visually.
• Exaggeration — widen a critical road so it remains visible.
• Typification — replace 50 individual trees with a symbol representing "forest".

Modern web maps apply generalisation dynamically per zoom level, often using tile-time pre-processing with tools like tippecanoe or tegola.

The Douglas–Peucker algorithm in one paragraph

Given a polyline, pick the two endpoints and find the vertex furthest from the straight line between them. If that distance is less than a tolerance ε, discard every intermediate vertex. Otherwise recurse on the two halves. The algorithm produces a simpler line that's never more than ε from the original — perfect for showing rivers at small scales without losing topology.

1from shapely.geometry import LineString
2[object Object]
3[object Object]
4

Scale effects in statistics: MAUP

Aggregating data into zones changes statistical results — even regression coefficients can flip sign. This is the Modifiable Areal Unit Problem (MAUP), first formalised by Openshaw in 1984. It has two facets:

• Scale effect — the size of the zones matters.
• Zoning effect — even at a fixed size, how the zones are drawn matters.

Practical mitigation: report results at multiple scales, use fine-grained data when available, and be cautious with inferences from aggregated units.

The ecological fallacy

Related: drawing conclusions about individuals from aggregated data. "Counties with high incomes have high car ownership" does not mean "wealthy people own more cars" — it might, but county-level correlation doesn't prove individual behaviour. Keep unit-of-analysis explicit in every claim.

Visualising scale dependence

A useful practice: when you produce any map, also produce the same map at half and double scale. Patterns that appear only at one scale should be reported with caveats; patterns that persist across scales are more robust.

Self-check exercises

Summary

• Three "scales": cartographic ratio, ground resolution, zoom level — know which one you mean.
• Data must be finer than the question.
• Generalisation (selection, simplification, aggregation, displacement) adapts data across scales.
• MAUP and the ecological fallacy punish the unwary.

Further reading

• Openshaw, S. (1984) — The Modifiable Areal Unit Problem.
• Douglas, D. & Peucker, T. (1973) — Algorithms for the reduction of the number of points required to represent a digitized line.
• Slocum et al., Thematic Cartography and Geovisualization.
• Mapbox — Vector tiles and generalisation at scale.