Global Fields of The World (FTW)

Name

llms.txt

Size

2.26 KB

Content Type

text/markdown; charset=utf-8

Last Modified

30 Jul 2026

Source URL

https://data.source.coop/ftw/global-data/predictions/zarr/llms.txt

Cloud URI

s3://us-west-2.opendata.source.coop/tge-labs/ftw-global-data/predictions/zarr/llms.txt

# FTW Global — Field Prediction Probabilities (Zarr)

PRUE model (https://huggingface.co/wherobots/prue-pt2) per-pixel class probabilities
over the FTW global Sentinel-2 feature mosaics, as one cloud-native Zarr V3 datacube.
Part of Fields of the World; paper: https://aka.ms/ftw-global-paper.

## Store
- Location: https://data.source.coop/ftw/global-data/predictions/zarr/alpha/global.zarr
- Format: Zarr V3, CF-1.8, EPSG:4326 (WGS84), GeoZarr-style (`proj:`/`spatial:` conventions),
  multiscale. Media type: `application/vnd.zarr; version=3; profile=multiscales`.
- Dimensions `(time, band, y, x)` = `(2, 3, 1566049, 4007517)`, variable `variables`
  is float32 softmax probability, chunks `(1, 3, 8192, 8192)`, NaN fill.

## Dimensions
- `time` (2): 2024, 2025 — CF "days since 2024-01-01".
- `band` (3): `non_field_background`, `field`, `field_boundaries`.
- `y` (1,566,049): latitude, ~10 m (8.983119e-5°), 83.748345 → −56.9317.
- `x` (4,007,517): longitude, ~10 m, −180 → 180.

## Relationships
- Same grid as `features/zarr` (stackable: open both with xarray/rasterix).
- The `vectors` (polygons) and `confidence` (raster) collections are derived from these
  probabilities.

## Multiscales
The root group declares `multiscales` against the WGS84Quad tile matrix set, resampling
`average`, with 14 levels: full resolution at path "." (factor 1) and overviews at group
paths 2x, 4x, 8x, 16x, 32x, 64x, 128x, 256x, 512x, 1024x, 2048x, 4096x, 8192x (factors
2 ... 8192; 8192x is 192 x 490). Each level is a complete group with the same
(time, band, y, x) layout, its own coordinate arrays and `spatial:transform`. Open one
with `xr.open_zarr(store, group="64x")`. The store is therefore readable at any zoom and
tileable for web display, not analysis-only.

## Access (xarray)
    import xarray as xr, rasterix
    ds = xr.open_zarr(
        "https://data.source.coop/ftw/global-data/predictions/zarr/alpha/global.zarr"
    ).pipe(rasterix.assign_index)
    field_2024 = ds["variables"].sel(time="2024-01-01", band="field")

## Caveats
- ~10 m global float32 cube (≈37 trillion values) — always subset by time/band/bbox;
  never load whole. Probabilities are softmax (sum≈1 across the 3 bands per pixel).
- License: CC-BY-4.0 (Taylor Geospatial Institute; Microsoft AI for Good Research Lab).