Multimodal Geo Vector

Multimodal Geo Vector

Turn visual-model annotations into reproducible GIS artifacts. Keep the reference raster—not the preview or model—as the authority for CRS, transform, extent and pixel grid.

Infer safe defaults from short requests

• Treat a named site or center coordinate as an AOI seed. Choose an initial extent appropriate to the target scale, expand it when candidates touch the edge, and record the radius/bounds. Do not hide a fixed default radius in the script. Interpret “all” as all visible candidates inside the recorded AOI; expand until candidates no longer touch the outer review margin or document the explicit coverage limit. Ask only when whole-site versus sample-area coverage materially changes the result or cost.
• Treat GeoPackage/.gpkg as a local projected-vector deliverable. Preserve the source CRS when suitable; otherwise use a local projected/UTM CRS and record it.
• Infer ordinary bands, preview size, tiling and overlap from the source and target. Do not require the user to restate the workflow in a long prompt.
• Record the mapping unit before inference: individual object, continuous same-target region, or enclosing site. Visibly separable roads, buildings, water, bare gaps and other excluded surfaces become polygon holes when they lie inside a continuous region. Ask only if another interpretation would materially change the output.
• Never infer CRS or bounds from visual appearance.

Route the request

1. Classify the source:
   • local georeferenced raster → local_first;
   • GEE image/collection followed by local vectorization → hybrid;
   • unreferenced JPG/PNG → require explicit CRS and bounds or georeferencing;
   • screenshot without a matching georeferenced raster → visual exploration only; do not promise CRS vectors.
2. For a GEE source, also use the sibling easygee skill. Run its interaction and geospatial-method routers, resolve the project from user settings, verify the dataset, and plan a batch export when a local download is too large.
3. Read references/workflow-profiles.md for target-specific geometry, tiling, prompt and QA choices.
4. Read references/annotation-contract.md before accepting model JSON, merging tiles, or diagnosing CRS misalignment.

Check readiness

Run:

python scripts/check_environment.py --mode local

Use --mode gee when Earth Engine access is required. Do not install or authenticate before checking. Never print or store GEE credentials. For GEE authorization/project issues, defer to the sibling EasyGEE auth workflow.

Prepare imagery

Local raster

Run:

python scripts/prepare_imagery.py local --input <image.tif> --output-dir <prepared> --bands 1,2,3

For an unreferenced JPG/PNG, accept it only with explicit map information:

python scripts/prepare_imagery.py local --input <image.png> --output-dir <prepared> --crs <EPSG:code> --bounds <xmin,ymin,xmax,ymax>

Do not infer bounds or CRS from scene appearance.

GEE source

Resolve project, AOI seed, bands, dates, scale, CRS and scale factor. For boundary extraction, prefer one actual acquisition over a median composite. Select a recent clear scene using AOI-valid-pixel coverage and AOI-local cloud quality; collection footprint coverage or scene-level cloud metadata alone is not enough. Run:

python scripts/select_recent_gee_scene.py --project <project> --center <lon,lat> --radius-km <km>

Pass its selected_image_id to preparation so the exact acquisition remains traceable:

python scripts/prepare_imagery.py gee --project <project> --output-dir <prepared> --region <xmin,ymin,xmax,ymax> --image-id <selected-image-id> --bands <red,green,blue> --scale <metres> --crs <EPSG:code> --name <scene>

Use a composite only when the user requests one or no adequate single scene exists, and label that fallback. The selector must fail when its coverage or clarity thresholds are unmet; use --allow-best-available only as an explicit, reported fallback. Verify the prepared raster's valid_pixel_fraction; reject or reselect scenes with clipped/nodata AOI coverage.

Use the EasyGEE export planner and ee.batch.Export.image.* for large results. Do not silently coarsen scale or shrink AOI to bypass download limits.

The preparation script must produce:

• a CRS-preserving raster;
• a model-friendly PNG preview;
• metadata containing source, bands, dimensions, transform, bounds, CRS and GEE provenance where relevant.

Choose whole-scene or tiled inference

Use the whole preview only when targets remain visually separable. For small or dense targets, create overlapping georeferenced tiles:

python scripts/tile_imagery.py --input <prepared.tif> --output-dir <tiles> --tile-size 512 --overlap 128

Prefer tiling when a target would be smaller than roughly 8–15 pixels in the model preview. Keep context around border objects.

Obtain multimodal annotations

Generate a strict prompt:

python scripts/prompt_template.py --image <preview.png> --target "<target definition>" --geometry <polygon|bbox|point|line> --output <prompt.txt>

Use the model's visual capability to return structured pixel JSON. Prefer JSON over painted overlays. Require the model to:

• use the declared preview width/height and upper-left pixel origin;
• return object id, label, geometry type, confidence and coordinates;
• avoid guessing map coordinates;
• lower confidence or omit ambiguous targets;
• follow explicit inclusion/exclusion rules.

If the model cannot return coordinates, request a single bright-color outline overlay that preserves the exact extent and aspect ratio. Treat overlay color extraction as a fallback because image edits can alter pixels or produce inner and outer duplicate contours.

Merge tiled annotations

Save tile-local JSON files as <tile_id>.json or <tile_id>_annotations.json, then run:

python scripts/merge_tile_annotations.py --manifest <tiles_manifest.json> --annotations-dir <tile-json-dir> --output <merged_annotations.json> --iou-threshold 0.5

Report object counts before and after overlap deduplication.

Restore CRS and export vectors

For structured annotations, run:

python scripts/vectorize_annotations.py json --annotations <annotations.json> --raster <reference.tif> --output-stem <results/targets>

For a bright-color overlay, run:

python scripts/vectorize_annotations.py overlay --overlay <overlay.png> --raster <reference.tif> --output-stem <results/targets> --color cyan

Apply --min-area, --max-area, --min-circularity or --simplify only when the target definition and projected CRS make the thresholds meaningful. Do not impose field-specific shape filters on vehicles, crowns or generic objects.

Write separate polygon, line and point GeoPackages in the raster's native CRS, EPSG:4326 GeoJSON exchange files, a polygon-boundary line GeoPackage, and a manifest with counts and paths.

Render a boundary-only QA overlay before handoff:

python scripts/render_vector_qa.py --raster <reference.tif> --vector <results/targets_polygons.gpkg> --layer polygons --output <results/qa_overlay.png>

Validate and hand off

1. Confirm raster and native vectors have the expected CRS.
2. Confirm all geometries are valid and output bounds overlap the raster.
3. Load the boundary line layer above the source raster in QGIS; use transparent polygon fill during QA.
4. Inspect duplicates, border truncation, shadows, touching objects and class confusion.
5. Preserve raw annotation JSON, prompt, model/version, thresholds and source metadata.
6. Call results candidate annotations unless independent labels support precision/recall, IoU or count-accuracy claims.
7. Report the selected image id/date, native pixel size, effective preview scale, candidate count, filters, and expected boundary uncertainty.

Run the no-data smoke test after modifying scripts:

python scripts/smoke_test.py

Do not bundle user imagery, case-specific labels or demonstration outputs in this skill. Use synthetic temporary data for offline tests or a user-selected, small GEE AOI for live verification.