Depave Medellín

This map screens the El Poblado neighborhood of Medellín for surplus pavement and ranks where its removal would do the most good. It is an international pilot, built entirely from global open datasets rather than municipal GIS, so read it as a methodology demonstration with coarser inputs than our U.S. sites.

Data tier 4 of 4 — Global open data (transferability demonstration) Medellín runs entirely on global open datasets: 10-meter ESA WorldCover, 30-meter Copernicus elevation, and OpenStreetMap, with no local equity layer and H3 hexagons in place of census tracts. Read it as a coarse, indicative demonstration that the method travels beyond U.S. municipal GIS, not a precise local analysis. Tiers reflect input-data availability, not effort. Scores are relative within each city, so the tiers and the numbers are not comparable across cities, and reliability decreases from Tier 1 to Tier 4.

Why depaving matters

Pavement raises local surface temperatures, sheds rainfall as runoff instead of letting it soak in, and crowds out the tree canopy that would otherwise cool the street. Some pavement is essential. Roads and sidewalks carry the movement a city depends on. Other pavement is surplus, like oversized parking areas and service lots. The tool looks for where that surplus sits in El Poblado and which parts of the neighborhood carry the heaviest environmental burden.

How this pilot differs

Our Fort Lauderdale and Bridgeport analyses lean on city and state GIS: 1-meter aerial imagery, surveyed road widths, lidar terrain, local canopy layers, and a federal equity screen. None of that is available here. Medellín runs on global datasets at 10 to 30 meter resolution. This is the coarsest and most fallback-driven of all our city methods. Where a U.S. site classifies pavement pixel by pixel, Medellín infers it from a 10-meter land-cover product. Treat every number on this page as a screening estimate with wider uncertainty than the U.S. sites carry.

How we identify pavement

We start with ESA WorldCover, a global 10-meter land-cover map. Its "built-up" class (class 50) captures everything sealed and human-made, including rooftops, roads, and lots. To isolate pavement, we subtract Microsoft Global Buildings footprints, which removes the rooftops and leaves the ground-level paved surface. There is no machine-learning classifier and no aerial-imagery step here. The classify stage that Fort Lauderdale uses is switched off for this pilot.

To separate optional from essential pavement, we compare each piece against a "core" mask built from OpenStreetMap. Road centerlines are buffered to an estimated width by their class, sidewalks are buffered to a fixed width, and mapped parking polygons are added. Pavement that falls inside this mask is core. Pavement outside it is non-core, the pool of depave candidates. Because there are no surveyed road widths for Medellín, the core mask depends entirely on OSM completeness and on class-based width estimates.

How we identify priority areas

There is no census-tract equivalent in this pipeline for Medellín, so we aggregate need to a grid of 14 H3 hexagons (resolution 8, roughly 0.74 km² each) covering the El Poblado pilot area. These are uniform hexagons, not comunas or barrios. For each hexagon we compute four need scores, each normalized to a 0–1 scale:

Heat: mean summer land-surface temperature from a Landsat 9 thermal composite.
Stormwater flood risk: a topography-only proxy derived from the Copernicus GLO-30 digital elevation model. Where does water tend to pond, and where does terrain concentrate flow?
Canopy deficit: one minus the fraction of the hexagon covered by tree canopy (ESA WorldCover tree class 10).
Pavement burden: non-core pavement area divided by hexagon area.

We average the four scores with equal weight into a single composite and flag the top quartile (hexagons at or above the 75th percentile) as priority hexagons. There is no equity overlay for this pilot, because no disadvantaged-community dataset comparable to the U.S. screening tools exists in this pipeline.

Headline findings (approximate, latest pipeline run)

Approximately 1,034 acres of pavement across the El Poblado pilot area.
Of that, about 291 acres are core and about 744 acres (72%) are non-core, and therefore candidates for depaving.
Need is aggregated to 14 H3 hexagons rather than census units.
4 of 14 hexagons (the top quartile) flagged as priority on the composite needs score.
No equity overlay is available for this pilot, so no "need meets disadvantage" finding is reported.

Key caveats

Pre-screening only. This tool scopes where to look first. Choosing specific parcels is a separate step, and site-level decisions require ground-truthing, utility checks, ownership review, and community input.
Global data are coarse. Pavement comes from a 10-meter land-cover product minus building footprints, far coarser than the 1-meter classifier used at the U.S. sites. Narrow alleys, sidewalks, and small parking aisles below roughly 10 meters are under-resolved.
No equity layer exists for this pilot. The U.S. sites overlay a federal disadvantaged-community screen. No equivalent dataset is available for Medellín in this pipeline, so the priority hexagons reflect environmental need alone, with no equity dimension.
Coarse spatial units. Need is summarized over 14 uniform hexagons rather than neighborhood census units, so results are coarse-grained. With only 14 hexagons, the top-quartile cut yields just 4 priority units.
The stormwater layer is a topographic proxy. It ranks where rainfall is most likely to pond using only terrain from a 30-meter elevation model. It does not model rainfall depth, drainage pipes, tides, or groundwater, and it has not been validated against observed flooding. Read it as a screen for where to look first, not a flood forecast.
The core/non-core split carries extra uncertainty. The road and sidewalk buffers that build the core mask are labeled in feet but applied in a meters-based coordinate system, which widens the mask. The exact effect on the split is under review. Treat the core and non-core acres as approximate.

For the full technical methodology, including data sources, algorithms, and known limitations, see the detailed methodology.