Pillar 2 — Nowcasting Hazard Susceptibility

1. From state to susceptibility¶

Given the soil state from Pillar 1 (saturation $S_w$, water-table depth $d_{wt}$, effective stress, strength) and current forcing, the nowcast estimates the likelihood and severity of each hazard at the present time. “Susceptibility” here is a probability of failure, not a binary map — the quantity that downstream warning and the forecast build on. Two hazard tracks are in scope:

• (a) Landslides — advanced; modeled with Landlab — full equations, pipeline, and limits on the Landslide Model page (§2).

• (b) Liquefaction & ground failure — in development; overview in §3, full model on the Liquefaction Model page (§3).

2. Landslides¶

2.1 Three types, three physics¶

Following the Varnes classification update Hungr et al., 2014, we treat three distinct failure processes that share data but not governing physics — each will become its own subpage:

The core GAIA digital-twin landslide products — shallow and deep-seated susceptibility — are rainfall- and groundwater-driven and do not take burn severity as an input. Post-fire debris flows are a special, wildfire-conditioned case of landsliding: only that track adds a burn-severity layer and fire-reduced cohesion. The current modeling effort targets shallow-landslide probability; the same Landlab framework extends to the others by swapping the hydrology and stability closures (and, for post-fire debris flows, adding the fire terms).

2.2 Modeling with Landlab¶

Landlab is an open-source Python toolkit for building two-dimensional numerical models of Earth-surface processes Hobley et al., 2017Barnhart et al., 2020. Its defining feature — the one that matters most for us — is component-based model coupling on a shared grid: a library of interoperable components (flow routing, snowpack, ecohydrology, slope stability) that each read and write named fields (topographic__elevation, soil__saturated_hydraulic_conductivity, …) on one common raster grid. This lets us chain a full hillslope hydrology-to-stability pipeline instead of stitching together disconnected models — the intermediate fields (e.g. soil moisture) are explicit, inspectable, and validatable (§2.6).

The landslide engine is Landlab’s LandslideProbability component Strauch et al., 2018, a probabilistic infinite-slope model driven by a topographically-controlled wetness index.

2.3 The physics (the equations we solve)¶

Infinite-slope factor of safety. On a planar failure surface at slope angle $\theta$, the factor of safety is the ratio of resisting to driving stress. In the dimensionless form used by the Landlab component Strauch et al., 2018:

where $\tilde C = (C_r + C_s)/(\rho_s g D)$ is the combined root + soil cohesion normalized by the saturated soil weight, $D$ is soil depth, $\phi$ the internal friction angle, $\rho_s, \rho_w$ soil and water densities, and $w\in[0,1]$ is the relative wetness (saturated fraction of the soil column). Failure is expected when $FS<1$.

Relative wetness (the topographic control). Following steady-state shallow subsurface flow Beven & Kirkby, 1979Montgomery & Dietrich, 1994, wetness rises with recharge and contributing area and falls with transmissivity and slope:

with recharge $R$, soil transmissivity $T$, specific contributing area $a/b$ (upslope area per unit contour width, from flow routing), and slope $\theta$. This is the SHALSTAB/SINMAP family of models Montgomery & Dietrich, 1994Pack et al., 1998; the unsaturated extension via suction stress is given by Lu & Godt, 2008Iverson, 2000.

From factor of safety to probability. The inputs $R$, $T$, $\tilde C$, $\phi$ are uncertain. The component draws them from distributions and runs a Monte Carlo ensemble, yielding a distribution of $FS$. The nowcast susceptibility is the probability of failure

i.e. the fraction of the ensemble that fails — a continuous 0–1 field, not a yes/no map.

2.4 Data clarity: raw vs. static vs. dynamic vs. derived vs. label¶

The most common source of confusion in this pipeline is conflating what is measured, what is computed, what is predicted, and what is used to check the prediction. They are different objects with different uncertainty and different roles:

RAW INPUTS ──► DERIVED / INTERMEDIATE FIELDS ──► PREDICTION ──► checked against ──► LABELS
(measured)     (computed by the model)            (P_f)                             (independent truth)

• Raw inputs are measured/observed and ingested via DataHub (Pillar 1).

• Static inputs do not change over a run; dynamic inputs are time-varying forcing.

• Derived / intermediate fields are computed by the model — they are not data, even though some (soil moisture, SWE) can be independently validated (your colleagues’ term: calibration).

• Predictions are model outputs (the probability of failure).

• Labels are independent observations of actual landslides, used only to score the prediction — never as model inputs.

Several of these layers (DEM, soil saturation, water table, precipitation) are shared with other hazards. The full layer-by-layer catalog is the single, cross-hazard Data Inventory, which carries hazard-purpose icons (⛰️ 🔥 🏚️ 🌊) marking which layer serves which hazard.

2.5 The prediction pipeline¶

Landslide probability is produced by gaia-hazlab/landlab-debrisflow (the “MMP” multi-model-probability workflow), which chains the Landlab components exactly along the taxonomy above:

terrain (DEM, flow accumulation)
   → static_inputs (soil, vegetation, transmissivity, cohesion
                    [+ burn severity — post-fire debris flows only])
   → daily_forcing (PRISM ppt, tmin, tmax)
   → snow (rain/snow partition, SWE, melt)
   → ecohydrology (PET, soil moisture)
   → landslides (routed recharge → LandslideProbability → P_f)
   → exports (GeoTIFF / ASC)

As currently configured for the Stehekin/Pioneer events, this is the post-fire debris-flow workflow — hence the burn-severity layer and fire-reduced cohesion. The core shallow and deep-seated landslide susceptibility runs the same stability engine and hydrology without the burn-severity input. It uses the data prepared in Pillar 1 (fire-debrisflow-ml and the SOLUS/PRISM staging). Today it still reads hardcoded local paths for those inputs; aligning it with the DataHub Integration Guide is the integration task that connects Pillars 1 and 2.

2.6 Calibration and validation¶

Two distinct checks, often conflated:

• Calibration — verify the intermediate fields against observations: soil moisture and SWE against in-situ sensors and the Pillar-1 ground-truth networks. This tunes the hydrology before any landslide claim is made.

• Validation — score the prediction $P_f$ against independent landslide labels: Sentinel-1 InSAR / SAR detections Mondini et al., 2021Handwerger et al., 2022 and, for the post-wildfire niche, post-event DEM/lidar differencing masks Bernard et al., 2021. These labels never enter the model.

Metrics (ROC/precision-recall and Brier score for $P_f$, spatial IoU for mapped failures, alert lead time) are defined in HazEvalHub.

2.7 Repository naming (suggestions)¶

The current names obscure what the repos do and should be clarified (tracked in the DataHub Integration Guide):

• landlab-debrisflow actually computes shallow-landslide probability (its internal package is debris_landlab), not only debris flows → suggest gaia-model-landslide (the susceptibility model).

• fire-debrisflow-ml is data preparation → suggest gaia-dataprep-landslide.

• This makes the data-prep (Pillar 1) → model (Pillar 2) split explicit in the repo names.

3. Liquefaction & ground failure¶

Earthquake shaking can turn saturated, loose granular soils into a fluid-like state — liquefaction — driving settlement, lateral spreading, and ground failure (hazard page). The trigger is seismic, not meteorological, but susceptibility is set by the same Pillar-1 state the landslide models use: saturation and water-table depth.

GAIA builds its geospatial liquefaction model (GLM) digital twin on the mechanics-informed surrogate of Sanger, Geyin & Maurer Sanger et al., 2025Sanger & Maurer, 2026. Rather than regress liquefaction observations on proxy variables — the approach of the models it is benchmarked against Zhu et al., 2017Rashidian & Baise, 2020 — it learns to reproduce a CPT-based geotechnical analysis at locations with no CPT, and defers to subsurface data where subsurface data exists. It is trained once and precomputed globally, so an event query is arithmetic over a stored raster.

The dynamic opening. Groundwater depth is by far the model’s most influential input, and it is currently a static value fixed at training. Making the water table a live variable is the main line of future work and the direct interface to Pillar 1 — the step that turns a static hazard map into a twin that responds to season, drought, and sea-level rise.

Resolution constrains all of it: liquefaction is controlled by fine-scale contrasts, so coarse static layers ($V_{s30}$, geology, water table) smear hazard. Full treatment on the Liquefaction Model page; layers and sources in the Data Inventory.

4. Evaluation & metrics¶

See HazEvalHub: event-based detection (POD, FAR, CSI), spatial agreement (IoU/Dice), probabilistic calibration (Brier, reliability diagrams) for $P_f$, and alert lead time.

5. Open questions & roadmap¶

• Quantify how prior uncertainty in static soil layers propagates to $P_f$ (sensitivity study).

• Close the Pillar 1 → Pillar 2 data loop by migrating landlab-debrisflow onto DataHub.

• Liquefaction: retrain the surrogate with the water table as a live input, and couple it to the Pillar 1 groundwater product (see Liquefaction Model §6).

• Liquefaction: integrate a time-varying site term ($\kappa_0(t)$, $V_s(t)$) into the NSHM hazard input for return-period products.

References¶