Given a city and a budget, which cooling measures should you build, and where?
shadecast is an open, credential-free pipeline for answering that, and this page walks it end to
end on three cities: Ahmedabad, Lagos and Rio de Janeiro.
The goal set at the outset was an open benchmark for urban heat adaptation planning. This is not
that yet. What exists is the pipeline and a first set of results: every layer taken from public
data with no API key at any step, every number produced by a real physics run rather than a
learned approximation, and every hypothesis committed to version control before it was tested,
including one that missed. Section 8 sets out plainly what is still missing and what it would
take to close the gap.
Nothing here is confined to the three cities in the selector, or to the 32 in the corpus. Every source is globally tiled and needs no account, and the study area is chosen by scanning population density, so the only requirement is that somebody lives in the search window. Six cities outside the corpus were checked directly, across five continents plus Iceland at 64 degrees north and Fiji on the 180th meridian: every source returned a file for all six. What limits this work is not the data, it is the physics.
| Cost | One city | 32 cities | 1,000 cities |
|---|---|---|---|
| Bundle on disk | 5.3 to 8.0 MB | 170 to 260 MB | 5.3 to 8.0 GB |
| Build, credential-free | 30 to 72 s | 16 to 38 min | 8.4 to 20 h |
| Physics, one evaluation | 162 to 283 s | 1.4 to 2.5 h | 45 to 79 h |
Only the first column was run: nine bundles built at 1 km² and 1 m, from 5.3 to 8.0 MB and 30.4 to 71.7 s each, with the physics timed on the same grid. The other two columns scale that measurement linearly and have not been run. Builds parallelise freely because every source is an independent HTTP read, and the population cache amortises: nine cities cost 297 MB of GHS-POP tiles between them rather than each. Physics does not parallelise away. One search of 1,000 evaluations on a single city costs the same as one evaluation on 1,000 cities, and a benchmark needs both, which is the whole reason a learned surrogate is the next step.