The coastline paradox explained for GIS work

Problem statement

"How long is the coast?" has no answer. The measured length depends on the length of the ruler, and as the ruler shortens the measurement grows without converging — which is the coastline paradox, and it is not a curiosity. It means any published coastline length is a statement about a dataset and a generalisation level, not about the world.

The effect is large enough to notice in ordinary work. Measured on the Natural Earth 10 m coastline within a box around Great Britain, the length falls from 9,136 km at full detail to 3,742 km when simplified with a 100 km tolerance — 41% of the original.

This guide quantifies it on a real dataset, explains why it happens, and covers what to do instead of quoting a length.

Quick answer

Measure geodesically and always publish the source and generalisation with the number:

import numpy as np, geopandas as gpd
from pyproj import Geod

geod = Geod(ellps="WGS84")

def geodesic_length_km(gdf):
    total = 0.0
    for g in gdf.geometry:
        c = np.asarray(g.coords)
        if len(c) < 2:
            continue
        _, _, d = geod.inv(c[:-1, 0], c[:-1, 1], c[1:, 0], c[1:, 1])
        total += float(np.nansum(d))
    return total / 1000

The same coastline at seven generalisation levels:

Length depends on the ruler, so measure geodesically

1. Stop treating length as a property of the coast

It is a property of the coast and the measurement scale. Two datasets of the same coast at different generalisation will always disagree, and neither is wrong.

2. Measure geodesically, not in degrees

Summing degree distances and multiplying by 111.32 km treats a degree of longitude as a degree of latitude. At 54°N a degree of longitude is 65 km, not 111 km, so the naive figure is far too large — and the error depends on how much of the coast runs east–west.

3. Measure in a projected CRS only if the projection preserves length

Most do not, and none does in every direction. For a national coastline an equidistant conic or the national grid is adequate; for anything continental, use geodesic distances.

4. State the source and its scale with the number

"9,136 km, Natural Earth 1:10 m, geodesic" is a reproducible statement. "9,136 km" is not.

5. Prefer scale-stable quantities

Area is far more stable than perimeter under generalisation, because simplification removes indentations that cancel. So are counts of features, distances between fixed points, and areas within a distance of the coast. If a metric can be expressed without a perimeter, express it that way.

6. If you must compare lengths, generalise both to the same tolerance

Two coastlines from different sources are only comparable after both have been simplified with the same tolerance — and that comparison measures change, not length.

7. Use the fractal dimension if the roughness itself is the point

The slope of log(length) against log(ruler) is 1 − D, where D is the fractal dimension. For coastlines D is typically 1.1 to 1.3, and it is a genuine descriptor of coastal character — much more meaningful than any single length.

Sweeping the ruler and estimating fractal dimension

The data used on this page

Natural Earth's 1:10m coastline and land polygons, clipped to a box around Great Britain (6.5°W–2.0°E, 49.8°–58.8°N).

import os
import subprocess

for name in ("ne_10m_coastline", "ne_10m_land"):
    if not os.path.exists(f"{name}.zip"):
        subprocess.run(["curl", "-sfLO", f"https://naciscdn.org/naturalearth/10m/physical/{name}.zip"], check=True)
print({name: os.path.getsize(f"{name}.zip") for name in ("ne_10m_coastline", "ne_10m_land")})

{'ne_10m_coastline': 3069451, 'ne_10m_land': 3269070}

Example 1 — the full sweep on your own data

import numpy as np, geopandas as gpd
from shapely.geometry import box

coast = gpd.read_file("ne_10m_coastline.zip")
gb = gpd.clip(coast, box(-6.5, 49.8, 2.0, 58.8))
gb = gb[gb.geom_type == "LineString"]

base = geodesic_length_km(gb)
print(f"{len(gb)} parts, {sum(len(g.coords) for g in gb.geometry):,} vertices, "
      f"{base:,.0f} km")

tolerances, lengths = [], []
for tol_km in (1, 2, 5, 10, 25, 50, 100):
    simp = gb.copy()
    simp["geometry"] = gb.geometry.simplify(tol_km / 111.32, preserve_topology=False)
    simp = simp[~simp.geometry.is_empty]
    L = geodesic_length_km(simp)
    v = sum(len(g.coords) for g in simp.geometry if g.geom_type == "LineString")
    tolerances.append(tol_km)
    lengths.append(L)
    print(f"{tol_km:4d} km ruler: {L:8,.0f} km  {L/base:6.1%}  {v:6,} vertices")

30 parts, 4,794 vertices, 9,136 km
   1 km ruler:    8,915 km   97.6%   1,668 vertices
   2 km ruler:    8,635 km   94.5%   1,006 vertices
   5 km ruler:    7,961 km   87.1%     490 vertices
  10 km ruler:    7,264 km   79.5%     304 vertices
  25 km ruler:    5,674 km   62.1%     141 vertices
  50 km ruler:    4,702 km   51.5%      97 vertices
 100 km ruler:    3,742 km   41.0%      75 vertices

The tolerance is in degrees — tol_km / 111.32 — so a "10 km ruler" is 10 km north–south but only about 6 km east–west at this latitude. That is good enough to show the effect, and a reason the rulers are approximate.

Example 2 — the fractal dimension

import numpy as np

tolerances, lengths = np.array(tolerances), np.array(lengths)    # from Example 1

slope, intercept = np.polyfit(np.log(tolerances), np.log(lengths), 1)
print(f"log-log slope {slope:.4f} → fractal dimension D = {1 - slope:.3f}")

log-log slope -0.1895 → fractal dimension D = 1.189

A dimension of 1.19 is typical of a moderately indented coast. A smooth sandy coast approaches 1.0; a deeply fjorded one approaches 1.3.

Example 3 — quantities that do not depend on the ruler

import geopandas as gpd, numpy as np

land = gpd.read_file("ne_10m_land.zip")
gb = land.clip(box(-6.5, 49.8, 2.0, 58.8)).to_crs(27700)       # clip in degrees, then project

for tol_m in (0, 1000, 5000, 25000):
    g = gb.geometry if tol_m == 0 else gb.geometry.simplify(tol_m)
    print(f"tolerance {tol_m:6,} m: area {g.area.sum()/1e6:12,.0f} km²  "
          f"perimeter {g.length.sum()/1000:9,.0f} km")

tolerance      0 m: area      240,739 km²  perimeter    10,929 km
tolerance  1,000 m: area      240,563 km²  perimeter    10,585 km
tolerance  5,000 m: area      238,328 km²  perimeter     9,382 km
tolerance 25,000 m: area      221,287 km²  perimeter     7,100 km

On this extract the area moved 0.1% at a 1 km tolerance and 1.0% at 5 km, while the perimeter lost 3% and 14%; at 25 km the area was down 8.1% and the perimeter 35%. Area is not immune — heavy simplification drops small islands whole — but across the tolerances anyone would use for analysis it is an order of magnitude steadier. That asymmetry is the practical reason to build coastal indicators on area rather than on length.

Why length never converges and area does

Why the length does not converge

A smooth curve has a length: as the ruler shortens, the polygonal approximation converges on it. A coastline is statistically self-similar over a wide range of scales — bays contain headlands which contain coves which contain boulders — so each shorter ruler finds new detail at roughly the same rate, and the total grows as a power of the ruler length rather than approaching a limit.

Why area is stable and perimeter is not

Simplification cuts across an indentation, removing a small area on one side and adding a small area on the other; the two largely cancel. It removes the entire boundary of that indentation from the perimeter, with nothing to cancel it. The result is that area converges quickly under generalisation and perimeter does not — which makes every ratio involving perimeter, including compactness indices, scale-dependent.

Why the effect is worse for some coasts

The fractal dimension measures how fast new detail appears. Norway's fjord coast has a much higher dimension than the Namibian coast, so the same change of ruler changes the Norwegian figure far more. Published national coastline lengths differ by factors of two to five between sources for exactly this reason, and the disagreement is entirely about scale rather than about the coast.

Why this matters beyond trivia

Anything normalised by coastline length inherits the problem: pollution per kilometre of coast, defences per kilometre, erosion rates per kilometre. Two regions measured from different source scales cannot be compared, and a time series that switches source mid-way records a change that did not happen.

Tolerances, topology and islands

• simplify tolerance is in CRS units. In degrees it is not a distance.
• preserve_topology=True limits simplification and changes the curve.
• Islands dominate the count. A coast's length is mostly its islands at fine scales.
• Tidal datum changes the line too. That is a separate, additive effect.
• Vertex density is not scale. A dense line can be a generalised shape.
• Raster-derived coastlines are stepped. Their length is a function of cell size.
• Publish the fractal dimension if roughness is the point.
• Never compare lengths across sources. Compare areas, or generalise both first.

Internal links

• Tidal datums explained: which shoreline is the shoreline — the other reason a coastline is ambiguous
• How to extract a coastline from a raster in Python — where cell size sets the scale
• How to measure shoreline change between two dates — comparing lines without comparing lengths
• How to simplify geometry in GeoPandas — the generalisation being applied
• Map scale and detail explained — generalisation as a cartographic decision
• How to measure distance accurately in Python — geodesic versus projected length
• Zoom and generalisation explained — the same effect in web maps
• Bathymetry explained: depths, datums and grids — resolution versus accuracy offshore

FAQ

Why does the coastline length change when I simplify?

Because a coastline is statistically self-similar: every shorter ruler finds new indentations. The measured length grows as a power of the ruler rather than converging.

How much does it actually change?

On a real Great Britain extract, from 9,136 km at full 1:10 m detail to 3,742 km at a 100 km tolerance — 41% of the original.

Should I measure in a projected CRS or geodesically?

Geodesically for anything larger than a region. Degree distances multiplied by 111.32 km are badly wrong wherever the coast runs east–west.

What should I publish instead of a length?

The length with its source and generalisation, or a scale-stable alternative such as area, or the fractal dimension if roughness is what you mean.

Why is area more reliable than perimeter?

Simplification removes and adds roughly equal areas on either side of an indentation, but removes its entire boundary from the perimeter.

Can I compare two countries' coastline lengths?

Only if both come from the same source at the same scale. Published national figures differ by factors of two to five between sources.