Geospatial

Base guardrail for all spatial work. Every spatial operation must pass these checks before methodology or tool-specific guidance applies.

Non-Negotiable Guardrails

These prevent silent corruption of spatial results. Check before any spatial operation.

1. CRS is mandatory — every geometry needs an explicit CRS (SRID). Without it, all spatial operations are meaningless.
2. No metric calculations on WGS84 — EPSG:4326 uses angular units (degrees). Always transform to an appropriate projected CRS before distance, area, or buffer calculations.
3. Use spatial index for joins — sindex (R-tree) on GeoDataFrames. Without it, spatial joins are O(n²).
4. Coordinate precision — 6–7 decimal places suffice for sub-meter accuracy. Beyond that is noise.
5. Ring orientation — exterior rings CCW, interior rings (holes) CW. Use shapely.geometry.polygon.orient() to normalize.
6. Axis order — GeoJSON is (lon, lat); Shapely is (x, y). Do not confuse them.
7. Null ≠ empty — None vs Point() have different semantics. None means unknown; empty means known but nothing there.
8. Common CRS before any operation — reproject all datasets to the same CRS before spatial joins, intersections, or any multi-dataset operation.
9. Nodata handling — raster Nodata values must be explicitly set and masked. Never treat Nodata as zero.
10. Alignment before raster math — rasters must share CRS, resolution, extent, and rotation before element-wise operations. Use rasterio.warp.reproject or xr.reproject.
11. Provenance — document data source, access date, processing steps, and CRS for every derived dataset.

Pre-Operation Checklist

When opening a new spatial dataset, always verify:

• CRS present and correct?
• Null or empty geometries? (filter or flag)
• Invalid geometries? (self-intersections, ring orientation — use geopandas.GeoSeries.is_valid)
• CRS mismatches across datasets?
• For rasters: Nodata value, band count, dtype, resolution?

Tool Selection

Use references/tools.md for curated library/tool selection, references/data-sources.md for dataset/source selection, and references/mcp-servers.md for MCP server selection. Load this skill for guardrails and routing; load the catalogs when choosing libraries, data sources, or MCP servers.

Format Decision Tree

Data type?
├── Vector (points, lines, polygons)
│   ├── Production / exchange → GeoPackage
│   ├── Cloud-native analytics → GeoParquet
│   └── Web tiles / quick sharing → PMTiles
├── Raster (2D grids)
│   ├── Single scene, web-ready → Cloud Optimized GeoTIFF (COG)
│   ├── Deep time series / datacube → Zarr / GeoZarr
│   └── Stitched large mosaic → COG with overviews
├── Point cloud (LiDAR, photogrammetry)
│   └── Cloud access → COPC (LAZ + embedded octree)
└── Metadata layer over any format
    └── STAC (SpatioTemporal Asset Catalog)

Task-Signal → Reference Routing

Load references on demand as the task requires. Use Context7 or official docs for any library API at use time — not this skill.

Tool selection and method reference may both be needed — load both when appropriate.

Cloud-Native Streaming Pattern

EO / ML pipelines should stream data instead of generating intermediate files:

STAC Catalog
  → pystac-client (discovery: bbox × time × cloud cover)
  → odc-stac / stackstac (lazy array: metadata only, no pixels loaded)
  → Dask (distributed: cloud masking, compositing, reprojection)
  → xbatcher (N‑dim batches → PyTorch/TF DataLoader)
  → GPU training

When to Use GEE vs Local Processing

When the Guardrails Are Violated

If you discover CRS mismatch, missing projection, invalid geometry, or unhandled Nodata — surface it as a Major finding. Do not proceed with analysis on spatially invalid data without explicitly noting the gap and its impact on result trustworthiness.