FTW Global — Field Prediction Probabilities (Zarr)

Per-pixel class probabilities from the PRUE model (wherobots/prue-pt2), run over the FTW global Sentinel-2 feature mosaics, as a single cloud-native Zarr V3 datacube. Part of Fields of the World; accompanying paper: https://aka.ms/ftw-global-paper.

The prediction Zarr opened in xarray (xr.open_zarr) — dimensions, bands, and chunking.

What's here

A single store: …/predictions/zarr/alpha/global.zarr, with dimensions (time, band, y, x):

• variables — float32 softmax probability per class, shape (2, 3, 1566049, 4007517), chunked (1, 3, 8192, 8192), NaN fill. CRS EPSG:4326 (WGS84), CF-1.8.
• Shares the exact grid of the features/zarr mosaic, so features and predictions are directly stackable.
• The vectors (field-boundary polygons) and confidence (raster) collections are derived from these probabilities.

Multiscales

This is a GeoZarr-style store (CF-1.8 plus the emerging Zarr proj: / spatial: geo-conventions) and it implements multiscales, so it can be read at any zoom and tiled for web display as well as consumed analytically.

The root group declares a multiscales attribute against the WGS84Quad tile matrix set, with 14 average-resampled levels. Full resolution lives at the root (path: "."); each overview is a sibling group named by its decimation factor:

Every level is a complete group — same (time, band, y, x) layout, its own coordinate arrays and spatial:transform. Open an overview directly by pointing xr.open_zarr at its group:

Using the data

Features and predictions can be opened side-by-side (same grid) to inspect inputs vs. outputs.

License

CC-BY-4.0. Produced by the Taylor Geospatial Institute and the Microsoft AI for Good Research Lab.