Data-Assisted Modelling and Uncertainty Analysis of Wetting–Drying Induced Ground Movement
HydroSense-Slope is a reproducible synthetic-twin framework for investigating how rainfall-driven changes in near-surface soil-water conditions propagate into progressive ground movement.
The project combines reduced-order hydrological modelling, van Genuchten soil-water retention behaviour, inverse parameter estimation, sensitivity and identifiability analysis, structural model comparison, and end-to-end uncertainty propagation.
Scope: This repository is a synthetic numerical benchmarking study. It is not a field-calibrated landslide forecasting model. The hydrological component uses a reduced-order root-zone water-balance formulation rather than a full Richards-equation finite-element solver.
The project addresses three linked questions:
- Can hydraulic parameters be recovered from noisy soil-moisture and suction observations?
- Which parameters and model components most strongly control the simulated hydro-mechanical response?
- How does hydraulic and structural model uncertainty propagate into predicted cumulative ground movement?
flowchart LR
A[Rainfall and evaporation] --> B[Root-zone water balance]
B --> C[Volumetric water content]
C --> D[van Genuchten SWRC]
D --> E[Matric suction]
C --> F[Degree of saturation]
C --> G[Inverse calibration]
E --> G
G --> H[Sensitivity and identifiability]
G --> I[Global uncertainty propagation]
F --> J[Ground-movement model]
A --> J
J --> K[Movement-model comparison]
I --> L[Hydraulic uncertainty]
K --> M[Structural uncertainty]
L --> N[End-to-end ground-movement prediction]
M --> N
The complete workflow can be summarized as:
Climate forcing → hydrological response → inverse calibration → sensitivity and uncertainty → ground movement → structural uncertainty → end-to-end prediction
Effective saturation is calculated as
where:
-
$\theta$ = volumetric water content, -
$\theta_r$ = residual water content, -
$\theta_s$ = saturated water content.
Matric suction is evaluated using the inverse van Genuchten relationship:
with
The parameters
Daily soil-water content is updated using a conceptual water balance containing:
- rainfall infiltration,
- limited infiltration capacity,
- short-term rainfall memory,
- evapotranspiration,
- gravitational drainage, and
- volumetric storage within a representative near-surface soil layer.
The formulation is intentionally computationally lightweight so that thousands of parameter realizations can be evaluated efficiently during inverse modelling and uncertainty propagation.
It should not be interpreted as a numerical solution of Richards' equation.
Incremental movement is represented using combinations of:
- background displacement,
- saturation above a threshold,
- antecedent three-day rainfall, and
- positive wetting rate.
Alternative model structures are compared rather than assuming that one phenomenological formulation is uniquely correct.
A 180-day synthetic monitoring sequence was generated using intermittent rainfall events and repeated wetting–drying episodes.
The simulated monitoring variables include:
- rainfall,
- evaporation,
- infiltration,
- runoff,
- drainage,
- volumetric water content,
- degree of saturation,
- matric suction,
- pore-water pressure, and
- cumulative displacement.
Figure 1. Synthetic rainfall forcing and resulting soil-water, suction, and ground-movement responses.
The generated dataset exhibits the expected qualitative hydro-mechanical behaviour:
- rainfall increases near-surface water content,
- increasing water content reduces matric suction,
- drying periods restore suction,
- repeated wetting episodes contribute to progressive cumulative ground movement.
Pseudo-observations were sampled every three days from the synthetic reference state and perturbed using measurement noise.
The first 120 days were used for parameter calibration. The remaining 60 days were retained as an independent hold-out validation period.
| Parameter | Synthetic truth | Estimated | Relative error |
|---|---|---|---|
| Infiltration efficiency | 0.8000 | 0.8261 | +3.26% |
| Drainage coefficient | 0.0750 | 0.0807 | +7.65% |
| van Genuchten α | 0.0600 | 0.0588 | -2.06% |
| van Genuchten n | 1.6000 | 1.6397 | +2.48% |
The recovered parameters remain close to their synthetic reference values despite measurement noise and an intentionally biased initial guess.
| Dataset | VWC RMSE | Suction RMSE |
|---|---|---|
| Calibration | 0.00398 | 1.507 kPa |
| Validation | 0.00688 | 1.593 kPa |
Figure 2. Inverse hydrological calibration and independent validation of volumetric water content and matric suction.
The validation results show that the calibrated model retains predictive performance outside the calibration period.
Successful parameter estimation does not necessarily imply that all parameters are independently identifiable.
A local sensitivity analysis was therefore conducted around the calibrated solution.
| Parameter | Normalized local sensitivity |
|---|---|
| van Genuchten n | 26.464 |
| Infiltration efficiency | 19.153 |
| van Genuchten α | 10.252 |
| Drainage coefficient | 7.072 |
The parameter-correlation structure reveals strong trade-offs between selected parameter pairs, particularly between the van Genuchten parameters and between infiltration and drainage controls.
Figure 3. Local parameter sensitivity and parameter-correlation structure.
This analysis illustrates an important inverse-modelling distinction:
A model can reproduce observations accurately while individual parameter estimates remain correlated or non-unique.
Local sensitivity only characterizes behaviour close to the calibrated parameter set. A broader global analysis was therefore performed using Sobol sensitivity indices and Latin-hypercube parameter sampling.
The most influential parameter on mean volumetric water content was:
Infiltration efficiency, with a total-order Sobol index of approximately 0.749.
The most influential parameter on mean matric suction was:
van Genuchten n, with a total-order Sobol index of approximately 0.428.
Figure 4. Predictive uncertainty in soil-water content and matric suction together with global Sobol sensitivity indices.
The results show a physical separation between:
- parameters controlling the water-balance response, and
- parameters governing the transformation from water content to matric suction.
Several reduced-order movement formulations were calibrated using only the calibration period.
| Candidate model | Parameters | ΔAICc | Calibration RMSE | Validation RMSE |
|---|---|---|---|---|
| saturation_only | 2 | 0.00 | 0.0593 mm | 0.2079 mm |
| no_wetting | 3 | 1.24 | 0.0585 mm | 0.1666 mm |
| no_rainfall | 3 | 2.31 | 0.0593 mm | 0.2054 mm |
| full | 4 | 3.71 | 0.0585 mm | 0.1666 mm |
| rainfall_only | 2 | 53.12 | 0.1152 mm | 0.3508 mm |
| no_saturation | 3 | 55.46 | 0.1152 mm | 0.3508 mm |
| wetting_only | 2 | 122.10 | 0.2729 mm | 0.7584 mm |
The calibration-based AICc comparison favours a parsimonious saturation-controlled representation, although competing formulations retain non-negligible statistical support.
This indicates that movement mechanisms are not uniquely identifiable from the available synthetic observations.
The final analysis propagates uncertainty through the full modelling chain.
For each realization:
- hydraulic parameters are sampled;
- the hydrological response is recomputed;
- candidate movement models are recalibrated;
- AICc-based structural model weights are calculated;
- a movement-model structure is sampled;
- cumulative displacement is predicted; and
- synthetic observation uncertainty is included.
| Model structure | Selection frequency |
|---|---|
| saturation_only | 45.9% |
| no_wetting | 29.2% |
| no_rainfall | 16.3% |
| full | 8.7% |
The final scenario-based displacement range is:
- 5th percentile: 3.870 mm
- Median: 4.058 mm
- 95th percentile: 4.207 mm
The independent validation-period performance is:
- Median prediction RMSE: 0.1675 mm
- Hold-out coverage of the 5th–95th scenario envelope: 30.0%
Figure 5. End-to-end hydro-mechanical prediction incorporating hydraulic-parameter uncertainty, movement-parameter uncertainty, observation uncertainty, and movement-model structural uncertainty.
The 5th–95th scenario envelope covers only 30.0% of the independent hold-out observations. This under-dispersion indicates residual model-form or extrapolation discrepancy that is not explained by parameter uncertainty alone. The envelope is therefore reported as a scenario range rather than a calibrated probabilistic confidence interval.
The synthetic-twin experiment demonstrates that:
-
Hydraulic parameter inversion is feasible under noisy observations. The principal hydrological parameters were recovered with relatively small errors from sparse pseudo-sensor measurements.
-
Predictive accuracy and parameter identifiability are different problems. Strong parameter correlations remain even when volumetric water content and suction are reproduced accurately.
-
Different parameters control different parts of the modelling chain. Water-balance parameters primarily influence volumetric water content, whereas van Genuchten parameters strongly affect matric suction.
-
Ground movement is dominated by hydrological state in the reduced-order experiment. Saturation-based formulations received the strongest statistical support.
-
Ground-movement model structure is not uniquely determined. Several alternative formulations remain plausible, motivating explicit treatment of structural uncertainty.
-
End-to-end uncertainty propagation produces a scenario range rather than a single deterministic trajectory. The final displacement is communicated using percentile bounds instead of one nominal value.
-
Parameter uncertainty does not fully explain hold-out prediction bias. The final 5th–95th scenario envelope remains under-dispersed during the validation period, indicating residual reduced-order model-form or extrapolation error.
HydroSense-Slope demonstrates a workflow that combines:
- physics-informed numerical modelling,
- inverse parameter estimation,
- independent validation,
- sensitivity analysis,
- parameter identifiability assessment,
- global uncertainty quantification,
- model-selection uncertainty, and
- hydro-mechanical response prediction.
The framework is designed as a computational research prototype that can later be extended toward higher-fidelity coupled geo-hydro modelling.
The framework intentionally contains several simplifications.
- The hydrological model is not a full numerical solution of Richards' equation.
- Hydraulic hysteresis is not represented.
- The ground-movement formulation is phenomenological rather than constitutive.
- Elastoplastic soil behaviour is not explicitly modelled.
- Progressive stiffness degradation is not included.
- Strain localization is not represented.
- Surface erosion is not coupled to the displacement model.
- The observations are synthetic rather than field measurements.
- Hydraulic parameter uncertainty is scenario-based rather than a Bayesian posterior.
- The percentile envelope is scenario-based and is not a calibrated Bayesian or frequentist confidence interval.
- Hold-out coverage remains substantially below the nominal 90% percentile range, indicating residual movement-model discrepancy.
- The predictive results should not be interpreted as site-specific landslide hazard probabilities.
The results should therefore be interpreted as a methodological demonstration rather than a field-scale hazard forecast.
The current framework provides a foundation for future implementation of:
- transient unsaturated flow governed by Richards' equation,
- coupled hydro-mechanical finite-element analysis,
- suction-dependent elastoplastic constitutive behaviour,
- wetting–drying hydraulic hysteresis,
- cyclic stiffness degradation,
- progressive damage,
- erosion coupling,
- Bayesian parameter inference,
- surrogate modelling,
- InSAR displacement measurements,
- pore-pressure and soil-moisture sensor assimilation,
- distributed fibre-optic sensing, and
- multi-modal monitoring datasets.
hydrosense-slope/
├── data/
│ ├── raw/
│ └── processed/
│
├── results/
│ ├── figures/
│ │ └── final/
│ └── tables/
│
├── src/
│ ├── generate_dataset.py
│ ├── hydrology.py
│ ├── validate_hydrology.py
│ ├── generate_observations.py
│ ├── calibration.py
│ ├── identifiability.py
│ ├── uncertainty_analysis.py
│ ├── displacement.py
│ ├── generate_movement_observations.py
│ ├── calibrate_movement.py
│ ├── movement_diagnostics.py
│ ├── end_to_end_uncertainty.py
│ ├── movement_parameter_uncertainty.py
│ ├── publication_style.py
│ ├── build_final_figures.py
│ ├── plot_phase7_publication.py
│ ├── build_figure_contact_sheet.py
│ └── build_readme.py
│
├── run_pipeline.py
├── LICENSE
├── README.md
├── requirements.txt
└── .gitignore
Using Conda:
conda create -n hydrosense-slope python=3.12 -y
conda activate hydrosense-slopepip install -r requirements.txtThe main computational workflow consists of:
python src/generate_dataset.py
python src/validate_hydrology.py
python src/generate_observations.py
python src/calibration.py
python src/identifiability.py
python src/uncertainty_analysis.py
python src/generate_movement_observations.py
python src/calibrate_movement.py
python src/movement_diagnostics.py
python src/end_to_end_uncertainty.py
python src/movement_parameter_uncertainty.py
python src/build_final_figures.py
python src/plot_phase7_publication.pyThe complete analysis can also be reproduced from the project root with a single command:
python run_pipeline.py- Python
- NumPy
- pandas
- SciPy
- Matplotlib
- SALib
- nonlinear least-squares optimization
- Latin-hypercube sampling
- Sobol global sensitivity analysis
- AICc-based model comparison
- Monte Carlo uncertainty propagation
This project is released under the MIT License.
Research prototype complete.
HydroSense-Slope is intended as a transparent computational geotechnics and uncertainty-analysis portfolio project, not as a production landslide forecasting system.