Agent Guidance, Natural Earth Populated Places (1:50m)

Curated world city points that join to the countries Collection on ADM0_A3. Join by attribute, never spatially. 113 of the 1,251 points, Istanbul, Montevideo, and Geneva among them, fall outside their own country's generalized 1:110m polygon because coastal simplification pushes them offshore.

Query the GeoParquet data asset in place with DuckDB spatial, read_parquet('natural-earth-populated-places.parquet'), or load it with GeoPandas. It streams over HTTP range requests, so query the published URL directly rather than downloading first. For a quick preview use the thumbnail asset.

Quirks

• The United States capital is spelled with two spaces after the comma, so NAME = 'Washington, D.C.' matches nothing. Use LIKE 'Washington%' or ADM0CAP = 1 with ADM0_A3 = 'USA'.
• South Sudan is SSD here but SDS in the countries file. Remap one side before joining, or Juba, Malakal, and Wau drop out.
• The legacy LATITUDE and LONGITUDE columns diverge from the geometry by up to 0.22 degrees. They are old label anchors, read geom.
• The UN time series columns POP1950 through POP2050 are populated only for the 463 UN-tracked agglomerations, zero elsewhere, and POP2050 is a projection.
• 59 places carry POP_MAX below 1,000, mostly stations and outposts, so population filters silently drop real capitals of small territories.

Tested Queries

Capitals per continent, with the country join done correctly.

Nearest city to a point, distance in kilometres.

Related Collections

reference/natural-earth-countries is the polygon companion, same producer and vintage family, joined on ADM0_A3 with the South Sudan remap. Converted from the upstream zipped Shapefile with DuckDB spatial into web-optimized GeoParquet 2.0.