Directory
‹ Back to apps

TerraLab GeoAI Skills

Geospatial AI skills for coding agents: the right CRS and measurements, PyQGIS for QGIS 4, Sentinel-2 indices, open imagery and map data, geoprocessing order, imagery for object detection, map export checks.

What’s inside

8 skills
  • Skillcrs-and-measurementsPicks the right coordinate reference system (CRS) for a GIS task, including national systems (Lambert-93 EPSG:2154, French CC zones, UTM zones, British National Grid, Swiss, Dutch, German, Australian grids), handles datum shifts, vertical datums and axis order, and makes distances and areas correct. Use when the user asks which CRS or EPSG code to use, reprojects a layer, gets wrong areas, lengths or buffers, sees data in the wrong place (swapped lat/lon, offset by meters or hundreds of km), mixes heights from different vertical datums, or asks how to check a layer's projection in QGIS, GDAL (ogrinfo, gdalsrsinfo) or GeoPandas.
  • Skillgeoprocessing-checklistPuts vector and raster geoprocessing steps in the right order and checks each result. Covers CRS checks and reprojection, fixing invalid geometries, clip, buffer, dissolve, overlay and spatial join, raster alignment before map algebra, with QGIS Processing algorithm ids, qgis_process and processing.run, PostGIS and shapely equivalents. Use when someone asks to buffer, clip, intersect, dissolve, join by location, compute areas or distances, run map algebra, chain QGIS tools, or asks why a buffer, area or overlay result looks wrong.
  • Skillimagery-for-object-detectionPrepares aerial, drone or satellite imagery before running an object detection or segmentation model, then checks the polygons it returns. Covers the ground sampling distance each object needs (buildings, cars, trees, solar panels, swimming pools), why Sentinel-2 at 10 m cannot find cars or small buildings, tile size and overlap, RGB versus multispectral, reprojection and resampling pitfalls, and result QA (tile-edge duplicates, minimum area, simplify, merge, spot checks, precision and recall). Use when someone wants to detect, count, segment or extract buildings, cars, trees, solar panels or pools from imagery, asks which imagery or resolution to use, or gets bad or duplicated detections.
  • Skillmap-export-qaChecklist to run before exporting, printing or sharing a map. Covers the map frame CRS for the area, classification methods (natural breaks, quantiles, equal interval) and when each misleads, sequential versus diverging color ramps, color-blind safe palettes and ColorBrewer, normalising counts by area or population, scale bar correctness (not in Web Mercator at high latitudes), north arrow, legend, source and license credits, DPI for print and screen, and QGIS print layout tips. Use when someone asks to export, print, publish or review a map, make a choropleth or thematic map, pick colors or classes, add a legend or scale bar, or set up a QGIS print layout.
  • Skillopen-imagery-finderChooses open satellite imagery, aerial orthophotos and elevation data for a place and a date (Sentinel-2, Landsat 8 and 9, Sentinel-1 SAR, Copernicus DEM GLO-30 and GLO-90, NAIP, national orthophotos such as IGN BD ORTHO, ESA WorldCover) by resolution, revisit, coverage and license, then finds scenes through STAC (Earth Search, Microsoft Planetary Computer) with a cloud cover filter and loads them in QGIS as COG, XYZ or WMS. Use when the user asks where to download satellite images, which imagery or DEM to use, how to get a cloud-free Sentinel-2 or Landsat scene, how to search a STAC catalog with pystac-client, or how to open a COG or an orthophoto in QGIS.
  • Skillopen-map-dataChooses and downloads open vector map data for anywhere in the world, and states the license and attribution it needs. Covers OpenStreetMap (Overpass API, Geofabrik extracts), Overture Maps (buildings, places, transportation, divisions, addresses, base) as GeoParquet queried with DuckDB, national sources such as IGN BD TOPO in France and OS OpenMap Local in Great Britain, open building datasets from Microsoft and Google, licenses (ODbL, CDLA Permissive 2.0, CC BY 4.0), bbox queries, size limits and output formats. Use when someone needs building footprints, roads, points of interest, addresses, boundaries or land use for an area, asks where to get GIS data, how to download OSM or Overture data, or whether they may use it commercially.
  • Skillpyqgis-qgis4Writes PyQGIS code that runs on both QGIS 4 (Qt6) and QGIS 3.x, in the right execution context (Python console, standalone script, Processing script or plugin). Covers qgis.PyQt imports, fully scoped enums (Qt.ItemDataRole.UserRole, Qgis.MessageLevel.Critical, Qgis.WkbType, Qgis.GeometryType), exec() instead of exec_(), layer editing, extents, and finding processing.run parameters with processing.algorithmHelp. Use when the user asks for a PyQGIS script, a QGIS plugin, a Processing algorithm, migrating a plugin to QGIS 4 or Qt6, or fixing errors like "type object 'Qt' has no attribute", "no attribute exec_", "QgsWkbTypes", "iface is not defined" or a layer that does not refresh.
  • Skillsentinel2-indicesComputes correct surface reflectance and spectral indices from Sentinel-2. Handles the processing baseline 04.00 offset (BOA_ADD_OFFSET -1000 since 25 January 2022) without applying it twice, SCL cloud and shadow masking, L1C vs L2A, band numbers and resolutions, resampling, and the formulas for NDVI, NDWI (McFeeters), NDMI (Gao), NDBI, NBR and dNBR with sanity ranges. Use when the user computes NDVI or another index from Sentinel-2, gets NDVI values that look too low or shifted after 2022, masks clouds with the SCL band, mixes 10 m and 20 m bands, or works with Sentinel-2 in Earth Engine, Planetary Computer, Earth Search, Copernicus Data Space, rasterio or the QGIS Raster Calculator.