Global available-water-capacity map and derived data with an explicit area of applicability

Derived data products for PEDOFLUX Paper 1, a global benchmark of soil available- water-capacity (AWC) mapping. Against 17,696 laboratory water-retention measurements, a widely used global product matches measurements only weakly (R2 = 0.20-0.35), and machine-learning skill is strongly protocol-dependent: random cross-validation (R2 = 0.72-0.80) overstates fair, density-declustered spatial-block skill (0.47-0.50) by roughly 60%, while AWC extrapolates poorly beyond sampled continents. Issued with a reliability mask, the maps are trustworthy over only about one-third of land and 56% of cropland. This archive contains the derived data only; the full reproducible analysis code is on GitHub (see Related identifiers). Input datasets are public and are not redistributed. Contents (all rasters: GeoTIFF, EPSG:6933 EASE-Grid 2.0 Global, 1 km, 34735 x 14629, NoData -9999; reliable.tif is Byte, NoData 255): - product/awc_mean|lo|hi|width.tif Predicted AWC and its calibrated 90% prediction interval [m3/m3] - product/paws_mm.tif Plant-available water storage over the top 1 m [mm] - product/di.tif Dry-season climatic water deficit [mm] - product/reliable.tif Area-of-applicability mask (1 = reliable, 0 = outside) - tables/train_table.parquet, measured_table.parquet, global_awc_grid.parquet, esacci_sm_climatology_025.tif Derived training, validation and confrontation tables A full manifest with per-layer units is in README.txt inside the archive.