Pillar 1 — Soil Reanalysis Product

1. Why a “soil reanalysis”?¶

Atmospheric science has weather reanalysis — a continuously updated, physically consistent estimate of the state of the atmosphere that blends models with all available observations (e.g. ERA5-Land Muñoz-Sabater et al., 2021). Geohazard prediction needs the analogous product for the ground: a soil reanalysis that estimates the state of soils and shallow subsurface water at the space and time scales relevant to hazards and resource management — tens of meters and sub-daily, resolving the hillslope and the vadose-to-water-table column, not the 9–25 km surface layer that global products deliver.

The defining idea is to build a holistic, multi-perspective state of the soil’s hydromechanics, because two coupled subsystems jointly govern hazard susceptibility:

1. Water partitioning — how water infiltrates as soil moisture, drains to the groundwater table, and returns to the atmosphere through evaporation and capillary rise. This is the unsaturated-flow problem Richards, 1931 closed by a soil-water retention curve Genuchten, 1980.

2. Mechanical state — how water content and lithological structure control the elastic properties of a granular or fractured-rock model of the soil, and therefore its cohesion and strength. Increasing saturation lowers suction and effective stress Lu et al., 2010, reducing the strength that resists slope failure and liquefaction.

No existing product delivers both, coupled, at hazard scale. That gap is what Pillar 1 fills, and it is what distinguishes a soil reanalysis from a soil-moisture product. The mapping from observations to these state variables — turning time-lapse seismic velocity into water-table depth and saturation — is developed in Soil Hydromechanical Memory.

2. State variables¶

The product estimates, in space and time, the variables that downstream hazard and resource models use:

These feed the Hazards pages and the resource-management work in Groundwater & Soil Moisture. Landscape-evolution and debris-flow surrogates built on Landlab Hobley et al., 2017Barnhart et al., 2020, and the ground-failure / liquefaction models in ModelHub, each require a specific subset of these layers — defining that requirement list is a core DataHub task (§5).

3. Data to build the state of soils¶

The data are multi-scale, multi-source, and span static to highly dynamic in time. We organize them into a small taxonomy, and require every layer to carry an explicit provenance statement: data source · sensor/measurement · resolution · uncertainty.

3.1 Static soil & terrain properties¶

Type, texture, lithology, depth to bedrock, and topography. National-to-global digital soil-mapping products provide priors: POLARIS at 30 m Chaney et al., 2019, the USGS SOLUS100 soil layers (100 m, used today in our debris-flow data prep — §5), SoilGrids globally Hengl et al., 2017Poggio et al., 2021, and USDA SSURGO/gSSURGO. Topography comes from the USGS 10 m DEM. These set the parameter priors ($\phi$, retention parameters, $K_{sat}$) for the physics, but they are static and smooth meter-scale heterogeneity.

3.2 Water-related dynamic state¶

• Past rainfall — from regression of gauge records through to physics/AI rainfall models; the proximate trigger and the dominant control on antecedent wetness Guzzetti et al., 2008. Already staged through GAIA via PRISM and HRRR (see §5).

• Groundwater table — the saturated-zone boundary condition.

• Soil moisture — the vadose-zone state. Gridded estimates are available from reanalysis/regional models (e.g. CONUS404 SMOIS/TSLB, already pulled by gaia-data-downloaders), and gaia-cli already emits a standardized soil_moisture variable — but these inherit the coarse support and limitations of §3.3 and §4, which the reanalysis must reconcile against ground truth.

3.3 Two observational modalities (with honest limitations)¶

Ground-based sensors stream data at uneven cadence from very heterogeneous sources, but gaia-cli and agentic download give us good access. These are our ground truth, both climatological and hydrological:

• Hydromet networks — rain gauges, streamflow, in-situ soil-moisture and well/piezometer data (the kind aggregated by the International Soil Moisture Network Dorigo et al., 2021).

• Geophysical networks — a modality absent from every comparable product (§4):

  • surface strain / kinematics (GNSS, InSAR, tilt) — deformation of the ground;

  • subsurface mechanical change via time-lapse seismic velocity $dv/v$ (DAS and regional networks), which responds to pore pressure and saturation Sens-Schönfelder & Wegler, 2006Clements & Denolle, 2018Mao et al., 2022;

  • seismogenic signatures — the seismic fingerprints of the hazards themselves (landslides, debris flows).

Satellite imagery offers good spatial but sparse temporal resolution, and is so-so for soils. Several caveats must be handled explicitly:

• Footprint leakage. Agencies reprocess raw measurements into Level-2+ products, but a reprojected/retiled product inherits the native measurement footprint — a downscaled 9 km soil-moisture pixel (e.g. SMAP L4 Reichle et al., 2017, ESA CCI Dorigo et al., 2017Gruber et al., 2019) still has a 9–25 km support, and these artifacts leak into downstream products (e.g. hydrologic-framework regridding) unless explicitly tracked.

• Regionalization of proxies. Indices such as NDVI are not always appropriate in the Pacific Northwest because the proxy-to-state model (e.g. NDVI → soil moisture) is miscalibrated for this region.

• Clouds and snow. Optical imagery is cloud-contaminated (Sentinel de-clouding helps) and loses contrast over snow — limiting usefulness exactly when winter hazards peak.

4. State of the art — and the gap GAIA fills¶

Mature products exist for pieces of this problem, but none delivers a hazard-scale, hydro-mechanical soil state.

How GAIA differentiates — Pillar 1 is designed to complement, not duplicate, these:

1. Hydromechanical, not just hydrological. We estimate effective stress, stiffness, and strength — the variables that actually govern failure — by coupling water and mechanics (Soil Hydromechanical Memory). Comparable products stop at moisture.

2. Hazard-relevant resolution and depth. Tens of meters, sub-daily, resolving the vadose-to-water-table column rather than a 0–5 cm surface layer.

3. Geophysical in-situ constraint. Dense seismic/DAS $dv/v$ and strain give direct subsurface state observations — a modality no listed product uses.

4. Physically-informed fusion with explicit provenance. We fuse sparse ground truth with satellite imagery (§3.3) and track per-layer footprint and uncertainty, rather than letting downscaling artifacts leak downstream.

5. Operational and cloud-native. Delivered through DataHub/gaia-cli as a continuously updated state, not a static climatology — feeding the nowcast and forecast in real time.

6. Targeted where global products are weakest — the forested, snow-affected, steep PNW.

5. Homogenization and the GAIA DataHub¶

The scientific core of Pillar 1 is a homogenization scheme that fuses heterogeneous ground sensors with satellite imagery into improved space–time fields of the soil’s hydromechanical state — proper interpolation and gap-filling regularized by physical models, not statistical regridding (which propagates the footprint artifacts of §3.3).

GAIA already has the skeleton of a DataHub to build this on; Pillar 1’s job is to make the soil state a first-class product within it. The current architecture is a convention with three layers:

• Object store — a shared s3://cresst bucket (anonymous read, authenticated write) is the de-facto hub, holding COG/Zarr/Parquet under s3://cresst/{user}/.

• STAC catalogs — per-dataset static STAC catalogs on GitHub make data discoverable and odc.stac.load-able into xarray: solus-stac (SOLUS100 soil, 18 properties × 7 depths, 100 m), prism-stac / precip-stac (precipitation), and landlab-stac (derived 10 m soil + terrain + geotech model inputs).

• Staging & discovery — gaia-cli (gaia stage prism|hrrr|synoptic|all -i AOI -s START -e END -o ZARR) clips, harmonizes, and writes a Zarr DataTree; the catalog web map is the human-facing index.

The gap Pillar 1 closes. Today gaia-cli and the STAC catalogs mostly cover precipitation/meteorology, while the soil layers the hazard models actually use are still produced by ad-hoc pipelines and only retroactively cataloged, or not at all. The same SOLUS100 layers appear twice — as hardcoded GCS URLs inside fire-debrisflow-ml / landlab-debrisflow (which also commit personal absolute paths like /mnt/c/Users/.../Downloads/...) and as a structured catalog in solus-stac — with no shared mechanism. Two soil vocabularies run in parallel (SOLUS100 at 100 m vs POLARIS at 30 m in landslide-digital-twin), and the liquefaction & ground-failure work in da-seis-groundfailure has no soil inputs wired yet. The Soil Reanalysis Product is precisely the layer that closes this gap.

Recommended DataHub work (steered by this use case):

• Promote the soil state to a first-class STAC product on s3://cresst, following the solus-stac/landlab-stac pattern, with per-layer provenance — source · measurement · resolution · uncertainty.

• Extend gaia-cli with a soil-state staging path (static SOLUS/POLARIS priors + dynamic CONUS404 SMOIS/TSLB + the seismic-derived $S_w$/$d_{wt}$ of Soil Hydromechanical Memory) alongside the existing precip stages.

• Reconcile the SOLUS100 vs POLARIS vocabularies and map them to the Landlab soil__* variable names the models expect.

• Migrate the ad-hoc data-prep repos onto this pattern and codify per-hazard variable requirement lists.

The concrete, repo-by-repo migration steps and the “DataHub-ready” checklist live in the DataHub Integration Guide — the guide the data-prep repositories can improve from.

6. Evaluation & metrics¶

Pillar 1 outputs are validated against the ground-truth networks of §3.3 — see HazEvalHub:

• State accuracy — RMSE/bias of $S_w$ and $d_{wt}$ against in-situ soil-moisture sensors, wells, and ET products;

• Dynamics — temporal correlation and lag of the storm response;

• Physical consistency — hydrostatic and mass-balance checks across the fused field;

• Footprint honesty — propagated, not hidden, uncertainty from satellite support scales.

7. Open questions & roadmap¶

• The data-assimilation scheme that blends $dv/v$, rainfall, and remote sensing into a single consistent state (the analog of atmospheric DA, which most listed products lack for land).

• Spatial upscaling from dense DAS cross-sections to regional reanalysis grids.

• Calibrating rock-physics closures (Dvorkin & Nur, 1996; see Soil Hydromechanical Memory) and retention parameters Genuchten, 1980 from sparse PNW boreholes.

• Defining the per-hazard soil-property requirement lists with the Landlab and ground-failure modeling groups.

References¶