dataset · Global

CoordBench

Hugging Face dataset taylor-geospatial/CoordBench

Claim this listing

Represent CoordBench? Send a claim for review so we can discuss listing updates. Claims do not automatically grant ownership or verification.

Data offered docs-verified: 40 normalized source tables from 13 source families, one config per original source file/table · Point samples with lon/lat (WGS84), optional unix-ms timestamp, official split where the source defines one, and source row id · Environmental and remote-sensing variables by config name, e.g. air_temp, worldclim_bio, bt_bioclim, bt_biomass, bt_landcover, soilgrids, ecoregions, country, dm_lcmap_lc, dm_lucas_lc, dm_openet_ensemble, usavars_elevation, cdc_places · Raw unmodified source files in a raw/ folder alongside data/
Coverage llm-inferred: Global per ENTRY regions; subset names also indicate US (usavars_*, cdc_places, dm_us_trees), Canada (dm_canada_crops_*), Africa and Ethiopia (dm_africa_crop_mask, dm_ethiopia_crops), and CONUS (pdfm_conus27)
Formats docs-verified: parquet
Access docs-verified: Public Hugging Face dataset (ENTRY auth null). No single licence: mixed per config, most are CC-BY-4.0, CC-BY-SA-4.0 or public domain, some differ. Check the per-config licence table before redistribution or commercial use.

Use cases llm-inferred: Compare location-encoder embeddings on soil property targets (soilgrids config) to test whether coordinate-only models can support ground-condition screening · Benchmark coordinate-based land-cover classification (dm_lcmap_lc, dm_glance, dm_lucas_lc) as a method test before applying embeddings to habitat baseline prototyping outside the UK · Assess regional-holdout generalisation of coordinate-based models before using them on unseen catchments or sites
Audience llm-inferred: GIS analysts · geospatial ML engineers and researchers · environmental data scientists

Constraints llm-inferred: Built for comparing coordinate-to-embedding models, not as a fit-for-purpose data product for FRA, drainage or BNG work · Mixed licences per config; the repo does not provide a universal licence · Some tables are joined, procedurally sampled from polygon/raster sources, or source-parsed, so values are not always raw source values · Nodata values (-999, -1, ocean index 0) and untransformed values are preserved in copied tables · There is one config per source table, not per label task, so prediction targets must be derived from columns · No UK-specific config appears in the subset list despite the UK consultancy focus · No stated spatial resolution, temporal coverage or update frequency for the dataset as a whole

Verified 2026-10-15 via HF API: exists; 1333 downloads/30d, 2 likes.

Metadata observed 14/10/2026 · research sweep: manual-2026-10-15 (Hugging Face datasets)