| English | 中文 |
Spatial analysis connects locations and supports, not just tables with coordinate columns. State the question, geometry, time/boundary version and acceptable location error before choosing operations. The spatial skill supplies separate R and Python functions, calls and checks.
| Task | Decision before calculation | R tools | Python tools |
|---|---|---|---|
| Spatial matching | Identity join, containment, intersection, overlap or nearest? What cardinality, boundary and distance rules apply? | sf::st_join, st_intersects, st_within, st_is_within_distance |
GeoPandas sjoin / sjoin_nearest, Shapely predicates |
| Coordinate conversion | Verified source datum/CRS, axis order, area of use and required accuracy | sf::st_transform; terra::project for rasters |
GeoPandas to_crs; pyproj Transformer; Rasterio warp |
| Mapping | Variable meaning/unit, support, shared colour limits, projection and missingness | ggplot2::geom_sf, grid/facet helpers; optional ggspatial |
GeoPandas/Matplotlib; reviewed interactive output |
| Geocoding | Address-to-location or reverse? Which source/provider, precision, privacy and permitted workload? | Local gazetteer; tidygeocoder for an explicitly selected service | Local gazetteer; geopy as a service client |
Geocoding does not prove administrative containment, and assigning a CRS does not transform coordinates. An attractive map does not validate matching or measurement.
Prefer stable verified identifiers when they represent the same entities and boundary version. Otherwise choose a geometric relation. A point exactly on a shared polygon border can have zero strict within matches or multiple intersects matches. Retain matched, unmatched and ambiguous states; do not select the first result or automatically snap it to the nearest region.
Nearest matching needs a study-based maximum distance and tie rule. A nearest object always exists in a nonempty reference layer, even when scientifically implausible. Record distance, units, candidates, coverage and unmatched fraction. GeoPandas nearest joins can return multiple equidistant matches; its distances are planar and unreliable in a geographic CRS. See sf predicates and GeoPandas nearest joins.
Check invalid/empty geometry, duplicate keys, many-to-many joins and boundary vintage before aggregation. Geometry repair can split objects or change area/type: preserve the source and inspect changes. For long/lat work, document the spherical/ellipsoidal engine; sf’s s2 guide explains why planar assumptions cannot be transferred automatically.
st_set_crs / set_crs declare what existing numbers mean; st_transform / to_crs calculate new coordinates. Unknown source CRS needs source verification, not a guess based on numeric range. Retain the original coordinates and datum, including coordinate epoch where required.
Distance and area require a suitable local projected or geodesic/equal-area approach. “Projected” alone does not guarantee metre units or low distortion. Web map display coordinates are not a general measurement CRS. With pyproj, explicitly use always_xy=True when the contract is longitude/latitude, and inspect transformation availability/accuracy. Missing grids or unsupported operations are reasons to review the transformation rather than silently accept an approximation. See sf transformation, GeoPandas CRS guidance and pyproj Transformer.
Chinese web-map coordinate products need their own source/provider convention. Do not label GCJ-02 or BD-09 as WGS84 and expect an ordinary EPSG reprojection to correct them. Review a provider’s documented coordinate service and terms, such as the AMap conversion interface; no conversion API is invoked by this skill.
A population count, concentration, density and category need different operations. Choose the target grid’s CRS, extent, origin, resolution and missing mask together; sharing only resolution does not establish alignment.
| Value meaning | A defensible direction | Required check |
|---|---|---|
| Category | Nearest/mode appropriate to the source | Codes remain valid; no invented fractional categories |
| Intensive continuous field | Declared interpolation or area-weighted mean | Units/support and missing coverage; do not sum concentrations |
| Cell totals | Conservative redistribution appropriate to the original support | Counts/mass before and after, including edge/mask losses |
| Exposure summary | Explicit area or population weights | Weight denominator, population support and excluded population |
The local dfproj pattern uses terra reprojection but its historical sum default is unsuitable as a general method. Specify the resampling rule for the quantity; a method named “sum” alone does not establish conservation after changing extent/masks. Use terra’s template-based projection, Rasterio reprojection and, for partial polygon cells, exactextractr. No raster adapter or conservation validation is implemented here.
Use the correct geometry support, declare value units and retain missing-data styling. Fix comparable colour limits/breaks for paired scenarios; report values outside the scale rather than silently clipping them. Point, polygon and rectangular grid layers are distinct representations. A basemap is a versioned input, not an implicit file or online download.
Reuse the member’s point/polygon and grid/facet patterns through generalized R helpers. Basemap inputs, extents, legends and export dimensions belong to task settings; there are no private-path defaults or global boundary objects. Exports specify physical size, DPI and units. Scale bars/north arrows must match the displayed projection and extent. See ggplot2 sf mapping and GeoPandas mapping.
Normalize addresses without discarding meaningful unit/locality information. Keep source IDs, provider/version, returned coordinate convention, candidate precision and manual-review state. A city centroid is not a building coordinate. Reverse-geocoded text is not proof of a person’s location.
Begin with an authorized local gazetteer when possible. Online services need deliberate provider selection and their current privacy, rate, cache and license conditions. geopy/tidygeocoder are clients, not a universal free data source. In particular, the public Nominatim policy restricts generic agent/platform integration, prohibits personal/confidential submissions and imposes workload conditions; it is not this skill’s default provider. Follow the operating guidance.
Spatial preparation can support data acquisition, prediction and impact assessment without replacing their scientific validation. Operational use is shared across 01, 03 and 04; stage 02 remains conceptual only. Source data stay versioned in external assets, non-secret settings in personal configuration, credentials in their protected mechanism, and results under external assets/results/<project_id>/. Sensitive locations never enter shared examples.