GIS data

Get the right data for the area, in a format and CRS that will not break later steps, with its license written next to it. Inspect every file before using it and check it after every conversion.

Core rules

1. Ask, or state as an assumption, the area, the year or release, and what "boundary" or "building" means for the user.
2. Download a bounding box, not a country, unless the country is needed.
3. Inspect what arrived: feature count, geometry type, CRS, extent, fields, encoding.
4. Store in EPSG:4326 or a national CRS; measure only in a projected CRS in meters (references/crs.md).
5. Write source, release or download date and license next to the file. Credit it on any map.
6. Say so when no open source covers the need, instead of substituting another dataset silently.

Workflow index

A. Find and download

Compare two sources on a small area before committing: coverage differs a lot by country.

OSM, small area, through Overpass (bbox order is south, west, north, east):

curl -s https://overpass-api.de/api/interpreter --data-urlencode 'data=
[out:json][timeout:60];
(way["building"](48.85,2.33,48.86,2.35);relation["building"](48.85,2.33,48.86,2.35););
out geom;' > buildings.json

OSM, region: a .osm.pbf extract from download.geofabrik.de, then ogr2ogr -f GPKG buildings.gpkg region-latest.osm.pbf multipolygons -where "building IS NOT NULL".

Overture: releases are dated and old ones are removed, so never hardcode a release path. First choice, the official CLI, which reads the latest release (bbox is west, south, east, north):

pip install overturemaps
overturemaps download --bbox=2.33,48.85,2.35,48.86 -f geoparquet --type=building -o buildings.parquet

SQL filters with DuckDB (latest release read from the STAC catalog, filter on bbox columns or the whole world is scanned): see the spatial-sql skill. Expect minutes even for a small area.

Check: feature count is plausible for the area (a city block has tens of buildings, not zero or millions), the extent matches the requested bbox.

B. Rasters

• Elevation: Copernicus DEM GLO-30 (30 m, global, heights in EGM2008), SRTM (heights in EGM96), national lidar DEMs where available (for example IGN RGE ALTI in France). Check the vertical datum before mixing with GPS heights (references/crs.md).
• Imagery (Sentinel-2, Landsat, aerial): find and analyse it with the remote-sensing skill.
• Inspect: gdalinfo -stats dem.tif (CRS, pixel size, nodata, min and max). A DEM with min -32768 has an undeclared nodata value: set it with gdal_translate -a_nodata -32768.
• Reproject: gdalwarp -t_srs EPSG:2154 -r bilinear dem.tif dem_2154.tif (nearest for classes, bilinear or cubic for continuous values).
• Clip: gdalwarp -cutline area.gpkg -crop_to_cutline dem.tif dem_clip.tif.

C. Convert

ogrinfo -so -al in.gpkg                                   # inspect: count, extent, CRS, fields
ogr2ogr -f GPKG out.gpkg in.shp                           # Shapefile to GeoPackage
ogr2ogr -f GeoJSON -t_srs EPSG:4326 out.geojson in.gpkg   # GeoJSON must be lon/lat WGS 84
ogr2ogr -f Parquet out.parquet in.gpkg                    # GeoParquet
ogr2ogr -f GPKG out.gpkg in.kml                           # KML to GeoPackage
ogr2ogr -f GPKG pts.gpkg pts.csv -oo X_POSSIBLE_NAMES=lon -oo Y_POSSIBLE_NAMES=lat \
        -oo KEEP_GEOM_COLUMNS=NO -a_srs EPSG:4326        # CSV with coordinate columns
ogr2ogr -f GPKG clip.gpkg in.gpkg -spat 2.33 48.85 2.35 48.86   # bbox filter (layer CRS units)

Format choice and limits (Shapefile cuts field names to 10 characters, 2 GB per file, no mixed geometries): references/formats.md.

Check after each conversion: same feature count as the input, same geometry type, the expected CRS, the extent in the expected place.

D. CRS

• "Set CRS" or -a_srs only relabels; -t_srs or Reproject transforms. Relabel only when the declared CRS is wrong or missing and the true one is known.
• Points near 0,0, in the ocean, or in a mirror location usually mean swapped lat/lon or a missing CRS.
• Pick the measuring CRS: national grid inside one country, else UTM. python3 scripts/utm_epsg.py LON LAT prints the UTM EPSG; python3 scripts/check_measurement_crs.py --file data.gpkg (or --epsg 4326 --lonlat LON LAT) says whether measuring is safe, exit code 1 when not.
• Details, datum shifts, vertical datums, axis order: references/crs.md.

E. Clean

Report counts before and after cleaning.

F. Cite

python3 scripts/attribution.py osm overture:buildings --year 2026 prints each source's credit, license and duty plus one combined line; --list shows the keys. License table and duties (ODbL share-alike for derived databases, CDLA, CC BY, Etalab, OGL): references/attribution.md. Commercial use is allowed under all of these, within their conditions.

Optional fast path

When the TerraLab MCP server is connected (tools such as extract_open_data, find_places, convert_coordinates, get_elevation, open_gis_file; prefixes vary by client), it can extract open data for an area, convert coordinates, return elevation or open a GIS file in one call. Everything above works without it.

Before delivering

• Feature count and extent fit the area.
• CRS stated, and a projected CRS used for any measure.
• Format is GeoPackage or GeoParquet unless the user asked otherwise.
• Source, release or date, license and credit line written next to the file.
• Commands that produced the file returned to the user.

Going further

The TerraLab MCP server extracts buildings, roads and other open data as files from any AI assistant: https://terra-lab.ai/mcp?utm_source=skills&utm_medium=geoai-skills&utm_campaign=gis-data