USDA NRCS gNATSGO Soil Properties

Soil properties for the United States and its territories, produced by the USDA Natural Resources Conservation Service (NRCS). This product takes the USDA NRCS gridded National Soil Survey Geographic Database (gNATSGO) — a relational GeoPackage plus regional map-unit rasters — and reformats it into a cloud-optimized, version-controlled Icechunk Zarr store, where soil properties can be sliced by location and depth, no database joins required.

Contents

One Icechunk repository at v0.1.0.icechunk/ with one group per source region, each on its native grid and CRS (no reprojection or mosaicking):

Each region group holds two independently openable subgroups (plus a diagnostics subgroup), all at the region's native resolution:

• CRS: coordinates are pixel centres (metres or degrees per the table above); a CF spatial_ref variable carries the WKT and GeoTransform, so rioxarray and GIS tools georeference the arrays directly.
• Storage layout: zarr v3 sharded arrays, Zstd-compressed, with (1, 128, 128) inner chunks (64 KiB raw float32) grouped into (1, 4096, 4096) shard objects (64 MiB raw). Inner chunks are addressed by range request, so a field-scale or point query fetches ~tens of KB per variable and depth: a 40×40-pixel window reads that shard's 16 KiB index plus 1–4 inner chunks, ~3–10 KiB compressed each (measured on puerto_rico).
• Metadata: every array's attributes record its source table and fields, units, missing-value semantics, and the exact aggregation algorithm and version that produced it.

Depth intervals

The depth_interval coordinate of soil_properties mixes six standard layers and five cumulative zones; auxiliary coordinates (depth_top_cm, depth_bottom_cm, depth_kind, to_reported_profile_depth) carry the same information numerically.

profile means the interval extends to the reported depth of the soil profile, which varies by soil component — not to a fixed depth.

Versioning

The store path carries the version; each USDA release is published as a new store path alongside the old ones, so a pinned path or tag always returns the same data. This table grows with each release:

Minor versions are new USDA data releases; a major version bump marks a breaking structural change (re-chunking, re-gridding, changed variable semantics). New variables may be added to the current store additively without a version change.

Included parameters

soil_properties — by depth interval

All float32, (depth_interval, y, x), NaN where no data are available.

The diagnostics subgroup carries the shared sand/silt/clay coverage pair (particle_size_contributing_thickness in cm, particle_size_contributing_component_percent in %) for filtering partially covered map units.

map_unit_properties — one value per pixel

Integer arrays keep their sentinel values when opened with xarray (no NaN masking / float upcasting); each sentinel's meaning is in the array's missing_value_semantics attribute.

Reading the data

Anonymous read access, no credentials needed:

Processing

gNATSGO does not ship these rasters. The source package contains one raster per region whose pixel values are map unit keys (mukey), plus relational tables describing each map unit: its components (soil series with a percent-of-map-unit weight) and their horizons (layers with top/bottom depths and measured properties). This product flattens that hierarchy into rasters in four steps:

1. Extract — the required tables (component, chorizon, chfrags, corestrictions, cointerp, chtexturegrp, mapunit, muaggatt) are streamed out of the GeoPackage; every source artifact is checksummed.
2. Aggregate to one value per map unit — per property, depth interval, and map unit:
   • Horizon properties (sand … ph): each horizon is clipped to the depth interval; horizons combine within a component as a thickness-weighted mean, and components combine as a component-percent-weighted mean. Missing values are never treated as zero — the value renormalizes over the data that exist, with the contributing thickness / component percent published as diagnostics. Sand, silt, and clay share one eligibility mask so they still sum to ~100% where coverage is complete. Exceptions: ksat takes the least transmissive (minimum) horizon of the dominant component; terrain and texture_class take the dominant component's representative value directly.
   • Valu1 parameters (aws, soc, NCCPI, root zone, droughty, pwsl1pomu, component sums): the 2026 release no longer includes USDA's precomputed "valu1" table, so these are reconstructed from the relational tables following the USDA column definitions (AWS as awc × thickness; SOC as a carbon stock corrected for rock fragments; root-zone depth from restriction layers with a 150 cm cap; the documented hydric/drainage criteria for wetland percent). Where the release provides comparators, the reconstruction matches: AWS agrees with muaggatt to within 0.05 mm and dominant-component slope matches exactly for 99.98% of map units.
3. Rasterize — each map unit's values are painted onto the region's native map-unit raster grid: a pure lookup with no resampling, so every pixel is exactly its map unit's value and mukey is a byte-for-byte copy of the source raster.
4. Validate — sampled pixels are compared back against the source raster and the derived tables before a release is tagged.

Each aggregation choice is recorded per-array in the aggregation_method / aggregation_algorithm_id attributes. Full documentation of the transformations:

• data-reference.md — the complete specification: aggregation algorithms (§5), the Valu1 reconstruction recipes and their verification status (§6), encoding (§7), and the pipeline (§8).
• future-variables.md — every candidate variable not yet included and the decisions each needs.

Processing code (MIT licensed): https://github.com/colinahill/usda_gnatsgo_data

Provenance & license

• Source: USDA NRCS gNATSGO, 2026-02-13 release (7.9 GB archive: 56.75 GB GeoPackage, SSURGO schema 2.3.3, plus 12 regional 30 m map-unit rasters, 25.2 billion cells). Retrieved from the USDA NRCS gNATSGO page.
• Audit artifacts: the source manifest (with SHA-256 checksums), validation reports, and per-property coverage tables are published under audit/2026-02-13/ in this product.
• Caveat: the Valu1-family arrays are a reconstruction, not USDA's own published table (which no longer exists for this release); see the processing repository for the verification status of each family.
• License: US Public Domain. USDA NRCS soil survey data carry no copyright restrictions and are free to redistribute; USDA NRCS asks for acknowledgement when the data are used. This README and the store metadata are provided under the same terms.
• Citation: Soil Survey Staff. Gridded National Soil Survey Geographic (gNATSGO) Database for the United States of America and the Territories, 2026 release. United States Department of Agriculture, Natural Resources Conservation Service.