Skip to main content

Only the North Held Its Green

Thirty years of satellite greenness for 1,110 cities, drawn on a night-dark ground with sand for loss and green for gain, and one region whose typical city ends where it began.

// Gribouille comes from the typst-render preamble (assets/typst/_preamble.typ),
// so this file does not import it.

// One row per city and year: the share of the city's area that stays green
// most of the year, read from satellite imagery. Rows without a city code are
// UN-Habitat's regional and global averages, not cities.
// Source: data/urban.csv (TidyTuesday 2026-09-22).
#let records = csv("data/urban.csv", row-type: dictionary)

// 2025 is published for too few cities to compare, so the span is the last
// year every city has.
#let first-year = "1990"
#let last-year = "2020"

#let cities = {
  let found = (:)
  for row in records {
    if row.cityCode == "NA" or row.year not in (first-year, last-year) { continue }
    let city = found.at(row.cityCode, default: (
      city: row.cityName,
      region: row.sdgRegion,
    ))
    city.insert("y" + row.year, float(row.averageShareOfGreenAreaInCityUrbanAreaPct))
    found.insert(row.cityCode, city)
  }
  found
    .values()
    .map(c => (..c, change: c.at("y" + last-year) - c.at("y" + first-year)))
}

// A region with a handful of cities has no swarm to draw, so it is left out
// and the caption says so.
#let min-cities = 10

// The catalogue's region names, shortened where they do not fit a row label.
#let display = (
  "Northern America and Europe": "North America and Europe",
  "Western Asia and Northern Africa": "West Asia and North Africa",
  "Central Asia and Southern Asia": "Central and South Asia",
  "Latin America and the Caribbean": "Latin America and Caribbean",
  "Eastern Asia and South-eastern Asia": "East and South-East Asia",
)

#let regions = {
  cities
    .map(c => c.region)
    .dedup()
    .map(name => {
      let mine = cities.filter(c => c.region == name)
      (
        region: name,
        n: mine.len(),
        median: median(mine.map(c => c.change)).y,
        lost: mine.filter(c => c.change < 0).len() / mine.len(),
      )
    })
    .filter(r => r.n >= min-cities)
    .sorted(key: r => r.median)
}
#let kept = regions.map(r => r.region)

#let drawn = cities.filter(c => c.region in kept)
#let lost-overall = drawn.filter(c => c.change < 0).len() / drawn.len()

// The claim the title makes: one region's typical city ends where it began,
// and every other region's typical city ends greyer.
#let north = regions.last()
#assert(
  north.region == "Northern America and Europe" and north.median > 0,
  message: "North America and Europe is no longer the only region whose median city gained",
)
#assert(
  regions.slice(0, -1).all(r => r.median < 0),
  message: "another region's median city now gains green area",
)

// Night from orbit: a near-black ground, with vegetation and bare soil in the
// colours a false-colour satellite image gives them. The figure keeps its
// colours on the light and the dark site theme alike.
#let night = rgb("#101613")
#let ink = rgb("#e6e9e2")
#let soil = rgb("#e3b061")
#let canopy = rgb("#3a9a8a")

// Barlow was drawn from the lettering of road signs and number plates, the
// type a city is read in. It is vendored in assets/fonts.
#let body-font = "Barlow"

// A change always carries its sign, with a true minus.
#let signed(v, digits: auto) = {
  let sign = if v > 0 { "+" } else if v < 0 { "\u{2212}" } else { "" }
  sign + format-number(digits: digits)(calc.abs(v))
}
#let share-label = format-number(digits: 0, suffix: "%")

// Rows are numbered from the bottom, so the region that lost most sits at the
// foot and North America and Europe at the top.
#let row-of = (:)
#for (i, r) in regions.enumerate() { row-of.insert(r.region, i + 1) }

// Each region is a swarm along its row, stacked here rather than with
// geom-beeswarm, which on a flipped or discrete axis draws each swarm under the
// wrong row label in this version of gribouille. Cities are binned by change, one
// percentage point to a bin, and the cities in a bin stack outwards from the
// row's centre line, alternately above and below, at the bin's centre. One step
// is shared by every row, so a taller stack means more cities, whichever region
// it is in.
#let bin-of(c) = calc.floor(c.change)

#let stacks = {
  let found = (:)
  for c in drawn {
    let key = c.region + "|" + str(bin-of(c))
    found.insert(key, found.at(key, default: ()) + (c,))
  }
  found
}
#let step = 0.8 / calc.max(..stacks.values().map(s => s.len()))

// A city is sand when it lost green area and green when it gained, in the same
// two colours the subtitle names. How much it changed is already its position.
#let points = {
  stacks
    .values()
    .map(stack => {
      stack
        .sorted(key: c => c.change)
        .enumerate()
        .map(((k, c)) => {
          let side = if calc.even(k) { 1 } else { -1 }
          (
            ..c,
            x: bin-of(c) + 0.5,
            y: row-of.at(c.region) + side * calc.quo(k + 1, 2) * step,
            paint: if c.change < 0 { soil } else { canopy },
          )
        })
    })
    .flatten()
}

// The median city of each region, as a bar across its row. The bar sits at the
// exact median, while each dot sits at the centre of its one-point bin.
#let bar-half = 0.46
#let medians = regions.map(r => (
  x: r.median,
  xend: r.median,
  y: row-of.at(r.region) - bar-half,
  yend: row-of.at(r.region) + bar-half,
))

// The panel reaches past the data on both sides: on the left for the row
// labels and on the right for the callout on the city that gained most. Both
// margins are a share of the data's span, so they keep their width in the
// figure if the data changes.
#let changes = drawn.map(c => c.change)
#let lo = calc.min(..changes)
#let hi = calc.max(..changes)
#let span = hi - lo
#let breaks = range(int(calc.ceil(lo / 20)) * 20, int(calc.ceil(hi)) + 20, step: 20)

// Each region is named on the left with its median and the share of its cities
// that lost ground, so the row reads without the axis.
#let names = regions.map(r => (
  x: lo - 0.045 * span,
  y: row-of.at(r.region),
  content: align(right, text(font: body-font, size: 7pt, fill: ink)[
    #text(weight: "semibold")[#display.at(r.region, default: r.region)] \
    #text(size: 6pt, fill: ink.transparentize(30%))[median #signed(r.median, digits: 1), #format-percent()(r.lost) of #r.n lost]
  ]),
))

// The two cities at either end of the range, named where they sit.
#let by-change = points.sorted(key: c => c.change)
#let callout(c, dx, dy) = (
  x: c.x + dx,
  y: c.y + dy,
  content: text(font: body-font, size: 6pt, fill: ink.transparentize(15%))[
    #text(weight: "semibold")[#c.city.split(" (").first()]
    #share-label(c.at("y" + first-year)) to #share-label(c.at("y" + last-year))
  ],
)

#plot(
  data: points,
  mapping: aes(x: "x", y: "y"),
  layers: (
    geom-vline(
      data: breaks.filter(b => b != 0).map(b => (b: b)),
      mapping: aes(xintercept: "b"),
      inherit-aes: false,
      colour: ink,
      alpha: 0.1,
      stroke: 0.4pt,
    ),
    geom-vline(xintercept: 0, colour: ink, alpha: 0.4, stroke: 0.6pt),
    geom-point(
      mapping: aes(colour: "paint", fill: "paint"),
      shape: "circle",
      size: 1.3pt,
      stroke: 0pt,
    ),
    geom-segment(
      data: medians,
      mapping: aes(xend: "xend", yend: "yend"),
      colour: ink,
      stroke: 1.4pt,
    ),
    geom-typst(
      data: names,
      mapping: aes(label: "content"),
      anchor: "east",
    ),
    geom-typst(
      data: (callout(by-change.first(), -0.6, 0.1),),
      mapping: aes(label: "content"),
      anchor: "south-west",
    ),
    geom-typst(
      data: (callout(by-change.last(), 0.8, 0),),
      mapping: aes(label: "content"),
      anchor: "west",
    ),
  ),
  scales: scales(
    x: scale-continuous(
      limits: (lo - 0.4 * span, hi + 0.27 * span),
      breaks: breaks,
      labels: signed,
      expand: (0%, 0%),
    ),
    y: scale-continuous(limits: (0.45, regions.len() + 0.55), breaks: (), expand: (0%, 0%)),
    colour: scale-identity(),
    fill: scale-identity(),
  ),
  guides: guides(default: none),
  labels: labels(
    title: "Only the North Held Its Green",
    subtitle: [
      Each dot is a city, placed by the change in the share of its area that is green from #first-year to #last-year, in percentage points. #text(fill: soil, weight: "semibold")[Sand] lost, #text(fill: canopy, weight: "semibold")[green] gained. \
      #format-percent()(lost-overall) of #format-comma()(drawn.len()) cities are greyer than they were. Only in North America and Europe does the median city end where it began.
    ],
    caption: [
      Green means vegetation that stays green most of the year in Landsat and Sentinel imagery; water is not counted. City boundaries follow the Degree of Urbanisation, not the municipality. \
      The white bar is each region's median city. 2025 is published for too few cities to compare. Regions with fewer than #min-cities cities are not drawn. \
      Source: UN-Habitat Urban Indicators Database (TidyTuesday 2026-09-22). Author: #link("https://mickael.canouil.fr")[Mickaël CANOUIL].
    ],
    x: none,
    y: none,
  ),
  theme: theme-minimal(
    ink: ink,
    paper: night,
    plot-title: element-text(font: body-font, size: 16pt, weight: "bold", colour: ink),
    plot-subtitle: element-text(font: body-font, size: 8pt, colour: ink),
    plot-caption: element-text(font: body-font, size: 6pt, colour: ink.transparentize(30%)),
    axis-text: element-text(font: body-font, size: 7pt, colour: ink.transparentize(20%)),
    panel-grid: element-blank(),
  ),
  width: auto,
  height: auto,
)

Figure 1: Only the North Held Its Green

Data: TidyTuesday 2026-09-22.

UN-Habitat reads greenness from space. For each city it takes satellite imagery, keeps the pixels that stay green for most of the year, and reports them as a share of the city’s area. Trees, shrubs and perennial grass count, and water does not. The series runs from 1990 to 2020 for every city, so the figure uses that span.

Across 1,110 cities, 65% have a smaller green share than they had in 1990. UN-Habitat’s own global average falls from about 15% to about 10% of city area over the same years.

The regions do not move together. In East and South-East Asia the median city lost 5.4 percentage points, and 79% of cities lost ground. Latin America and the Caribbean, Central and South Asia and Sub-Saharan Africa each lost about 2 to 4 points at the median. North America and Europe is the only region whose median city ends where it began, at plus 0.2 points, with its cities split almost evenly between gains and losses.

The extremes are in Africa and Asia. Kenema and Bo, in Sierra Leone, were three-quarters green in 1990 and are about an eighth green now. Chengdu fell from 69% to 15%. At the other end, Stoke-on-Trent went from 8% to 26%, the largest gain in the data.

Some cautions. A city here is the urban area defined by the Degree of Urbanisation, which can be larger or smaller than the municipality with the same name. The share says how much of that area is green, not whether parks were lost, so a fall can come from building over fields at the edge as much as from losing trees in the centre. The data for 2025 exists for only a small number of cities, so it is left out. The two cities in Oceania outside Australia and New Zealand are too few to draw as a row.