Potential-natural-vegetation world data

Potential vegetation is sourced from Martin Jung/IIASA, Current and future European potential vegetation types v1.1, DOI 10.5281/zenodo.14627466. The pinned version was published 2025-01-10 under CC BY 4.0 and covers continental Europe, including Turkey, at nominal 1 km grain. Canonical temporal metadata records the model-input window, 1990–2020, rather than presenting the publication year as a single observation year.

Run just init-jung-pnv to download the categorical current raster plus the six current-class COGs into target/world-data-sources/raw/jung-pnv/. The atomic initializer pins every official filename, byte size, published MD5, and verified SHA-256, rejects oversized streams, and writes a deterministic adjacent jung-pnv-manifest.json. Rasters and the manifest are ignored local source data; they are never distributed with the repository. Use --potential-vegetation-dir to select another verified initialization directory.

The importer validates the actual source contract: 5,583 by 4,474 Float32 cells, 1,000 m pixels, top-left (944000, 5416000), 512-square Deflate tiles, NaN nodata, and the source's user-defined GeoTIFF keys carrying the EPSG:3035 LAEA parameters. Posterior COGs contain seven interleaved samples named mean/sd/q05/q50/q95/mode/cv; world compilation consumes the mean and validates the complete band contract. The categorical raster accepts only 1 through 6. Every import revalidates the manifest's record/version/license/file identities, the exact byte sizes, and SHA-256 values before opening a raster. COG tiles are decoded on demand with checked channel lengths and a byte-accounted cache capped at 64 MiB.

Canonical cells receive nodata-aware area-weighted posterior means. When every class has a valid posterior, six independently quantized 0..10,000 basis-point scores are stored; they are not asserted to sum to 10,000. Otherwise the categorical raster is used, choosing greatest valid overlap with stable class ties. A cell with neither form of source evidence receives a deterministic non-unknown class inferred from already typed forest, elevation, latitude, and HYDE 3.5 context. Reports reconcile posterior, categorical, and inferred outcomes exactly to settlement count.

Inference-rules version 9 and world schema version 25 identify the complete post-hydrology synthesis contract. Older artifacts or caches cannot share identity with Jung-derived or reconstructed historical results.

Potential vegetation remains stored unchanged as the modern-climate ecological envelope. After soil and hydrology finalization, the compiler separately stores dominant 1544 cover. The greatest sampled HYDE 3.5 human fraction is selected first (stable tie order: built, cropland, pasture), then becomes direct only when it meets its own threshold: 10% built or 35% cropland/pasture. Deterministic missing-HYDE profiles are never labeled direct. Natural cover is derived from Jung, Copernicus forest structure, EU-Trees4F candidates, soil/geology, elevation, hydrology, latitude, and OWDA moisture. Only genuinely close natural scores use coordinate-and-schema hashing as a deterministic tie-break. Fallback Jung wetland/marine classes cannot emit water cover: wetlands require wet soil plus freshwater/tidal convergence, while transitional water requires tidal evidence.

Attribution/modification notice: Adventure Simulator downloads Jung's published v1.1 rasters unchanged, then projects settlement cells, area-aggregates posterior means, quantizes values, and applies documented categorical/inference fallbacks.

The terrain pipeline separately extracts a bounded wetland mask from the pinned 1 km wetland posterior (mean >= 0.5), using categorical class 5 only where the posterior is nodata. Pixel centers are clipped to playable bounds and capped at 100,000 candidates. The pinned SHA-256, accepted pixel count, and terrain package digest are recorded; this is source coverage, not settlement extrapolation.