Landslides

Landsliding spans a continuum of failure depths and timescales Hungr et al., 2014. The GAIA digital twin predicts a probability of failure $P_f$ for two end-members that bracket the core hazard; neither takes burn severity as an input.

1. Process & triggering mechanisms¶

Both inherit their state from Pillar 1 — Soil Reanalysis Product: shallow failures respond to the saturation field, deep failures to the water-table field.

2. Characteristics¶

3. Implementation in Landlab¶

The current implemented Landlab workflow uses the probabilistic infinite-slope LandslideProbability component to compute shallow landslide susceptibility from recharge-driven relative wetness and uncertain material properties Strauch et al., 2018. A deep-seated variant can be framed within the same stability concept, but should be treated here as a planned extension with a different hydrologic closure.

This shared framing is useful conceptually, but the current working GAIA notebook and packaged workflow are the shallow, recharge-driven implementation.

Data — what we ingest¶

The full, traceable catalog — every raw product, every deterministically derived layer, and every model output, with sources/APIs, access sensitivity, spatial/temporal resolution, and limitations — lives in the Data Inventory under the DataHub. In brief, the core (non-post-fire) model ingests:

• Terrain — a DEM (USGS 3DEP / OpenTopography) → topographic__elevation, slope, contributing area.

• Static soil properties — SOLUS100 (100 m) or POLARIS (30 m) → thickness, density, friction angle, cohesion bounds, $K_{sat}$/transmissivity, porosity, field capacity, wilting point.

• Vegetation — plant functional type / landcover used to assign vegetation parameters such as LAI.

• Forcing — observed PRISM precipitation/temperature (hindcast/daily) and Earth2Studio AI-weather precipitation (forecast).

• Reanalysis state — the saturation (shallow) and water-table (deep) fields from Pillar 1.

• Calibration targets — SMAP soil moisture, ERA5 SWE, and mapped landslide inventories.

No burn severity — that layer is specific to post-fire debris flows. See the inventory for sources, sensitivity, resolution, and the models behind each derived product.

Models¶

The full model documentation — physics equations, the data→Landlab pipeline, what is solved vs assumed, watershed/single-drainage constraints, Landlab limits for digital twins, Earth2Studio interoperability, and evaluation — is on the Landslide Model — Landlab Implementation page in the ModelHub. The implementation lives in gaia-hazlab/landlab-debrisflow; see also Pillar 2 §2.5.

Detection & monitoring¶

Susceptibility is the forecast side; GAIA also detects landslides as they happen. Rapid mass movements — debris flows, rockfalls, icefalls, lahars — radiate seismic and infrasound energy, so a continuously running detector can catalog events in near-real time, independent of the revisit gaps that limit optical/SAR mapping.

In the Mt. Rainier region this is done with QuakeXNet, a deep-learning detector run on continuous waveforms from seismic stations within ~50 km of the volcano: per-station detections are aggregated to network-level events, located with ENVELOC, and validated against the PNSN and ESEC catalogs — yielding a 15-year (2010–2025) catalog of ≈128,500 located events (≈115,000 surface events plus explosions) Kharita, 2025. The trained model is distributed through the SeisBench ecosystem; multi-sensor extensions (infrasound, tilt, DAS) and other PNW sites are in progress (see ModelHub). The catalog is currently under review, with ongoing refinement of the event-classification model, so event counts and class labels will continue to evolve.

These detections close the loop with the model in two ways: they provide validation labels for the modeled probability of failure $P_f$ — event presence, time, and location (see model §8 and HazEvalHub) — and they are a real-time monitoring stream alongside the nowcast.

Explore the located catalog on the interactive map (full project: pnw_seismic_event_detection):

Evaluation & metrics¶

(Link to HazEvalHub: probabilistic calibration (Brier, reliability) of $P_f$, spatial agreement (IoU) against detection masks and the seismic detection catalog, and — for deep-seated — displacement-rate skill and early-warning lead time.)

Connection to use cases¶

Shallow failures feature in the 2025 Western Washington floods & landslides.

Open questions & roadmap¶

• Quantify how prior uncertainty in static soil layers propagates to $P_f$ for each type.

• A unified hydrology closure spanning the shallow (vadose) and deep (water-table) regimes.

References¶