Landslide susceptibility and uncertainty mapping
A completed geology study comparing spatial and random validation while mapping slope-unit landslide susceptibility, model disagreement, bootstrap uncertainty, calibration, and residual spatial structure.

Software compatibility
The completed workflow uses the checksummed 7,360-unit Central Italy benchmark and the pinned Python 3.14 packages. Replacing the region, inventory, target, or factors requires a new provenance and spatial-validation review.
Project definition
Problem statement
Nearby slope units share terrain, geology, survey history, and inventory bias. Randomly mixing them between training and testing can overstate how well a susceptibility model transfers to a separated area.
A single probability map also hides disagreement among model structures and sensitivity to the training sample. The geology problem is to measure both spatial transfer and uncertainty without confusing susceptibility with hazard, risk, or site safety.
Project objectives
- Verify and align the public Central Italy slope-unit table and geometry with recorded checksums.
- Compare logistic regression, random forest, histogram gradient boosting, and an equal-weight ensemble.
- Use mutually exclusive spatial blocks as the primary estimate and matched random folds as an optimism diagnostic.
- Measure discrimination, probability quality, calibration, model disagreement, and bootstrap uncertainty.
- Test factor importance, inventory definition, lithology removal, block count, and residual spatial structure.
Project structure
Project components
Source validator
Checks file hashes, schema, identifiers, geometry alignment, target values, and the documented coordinate-reference conflict.
Spatial validator
Builds five coordinate-based blocks and guarantees that every retained prediction comes from a model that did not train on that block.
Model comparison
Fits one regularized linear model and two nonlinear tree families before combining their out-of-fold probabilities.
Uncertainty analysis
Separates model spread from 24-resample within-fold bootstrap variability and maps their combined magnitude.
Geological diagnostics
Calculates held-out factor importance, response profiles, inventory and feature sensitivity, and residual Moran autocorrelation.
Methodology
Project workflow
- 01Verify the source
The workflow checks the retained CSV and GeoPackage, aligns all 7,360 identifiers, and assigns the publication-supported EPSG:32632 definition.
- 02Build validation folds
Slope-unit centroids form geographically separated blocks while a matched random design supplies a direct comparison.
- 03Fit out of fold
Each model predicts only held-out units, and the three probabilities form an equal-weight ensemble.
- 04Measure uncertainty
Within-fold bootstrap distributions and cross-model spread produce intervals, total uncertainty, and a bivariate map.
- 05Interpret and test
Performance, calibration, factors, sensitivities, and residual spatial structure are retained as tables, maps, and report evidence.
Demonstration scenario
Across 7,360 slope units, the spatial ensemble reaches ROC AUC 0.7522, average precision 0.7201, and Brier score 0.2024. Random folds raise AUC to 0.7772, a 0.02494 optimism gap. The final screen separates 1,679 high-susceptibility lower-uncertainty units from 794 high-susceptibility higher-uncertainty units.
Engineering
Tools and method
- Tools
- The project uses Python 3.14, GeoPandas, scikit-learn, Pandas, NumPy, SciPy, Matplotlib for subject analysis, simulation, and results.
- Geospatial data layer
- GeoPandas validates and joins the slope-unit attributes, projected polygons, centroids, and retained GIS outputs.
- Model layer
- scikit-learn pipelines standardize the linear model, fit nonlinear alternatives, and generate leak-free out-of-fold probabilities.
- Uncertainty layer
- NumPy and SciPy combine model disagreement, bootstrap variability, calibration, concentration capture, and spatial residual testing.
- Evidence layer
- Pandas, CSV, JSON, GeoPackage, and PNG retain complete predictions, metrics, sensitivity results, maps, and figures.
- Verification
- Automated tests, branch coverage, dependency audit, Docker, document checks, provenance checks, and repository validation verify the delivery.
Testing
Evaluation
Evaluation measures
- ROC AUC, average precision, Brier score, log loss, and balanced accuracy
- Calibration intercept, calibration slope, and observed probability bins
- Random-minus-spatial performance gap under matched five-fold designs
- Model disagreement, bootstrap interval width, and total uncertainty by slope unit
- Spatial-holdout permutation importance and factor-response profiles
- Sensitivity to block count, lithology removal, and the stricter presence2 inventory
- Residual Moran I with 499 seeded permutations
Project boundaries
- The output is a static empirical susceptibility study for the retained Central Italy benchmark and inventory definitions.
- It contains no rainfall time series, pore-pressure model, seismic trigger, runout, exposure, vulnerability, or consequence estimate.
- Predictor importance describes model association under the source data and does not establish a complete mechanical cause.
- The map is not suitable for warning, construction clearance, evacuation, or slope certification.
- Transfer to an Indian or other region requires licensed regional data, field and institutional evidence, recalibration, and independent spatial validation.
Included
- 01Complete Python and GIS source code
- 02Retained public benchmark CSV and GeoPackage with checksums and provenance
- 03Out-of-fold probabilities and uncertainty for all 7,360 slope units
- 04Spatial and random five-fold validation with three model families and an ensemble
- 05Twenty-four within-fold bootstrap resamples and bivariate uncertainty mapping
- 06Six sensitivity scenarios and a 499-permutation residual Moran analysis
- 07Twenty analytical figures and two attributed literature images
- 08Sixty-seven automated tests with 97.50 percent branch-aware coverage
- 09Complete project files, calculations, results, and analysis material in a private GitHub repository
- 1098-page project documentation in PDF and editable Word formats
- 1110-page setup and usage guide in PDF and editable Word formats
- 12Fifty-two annotated references, including current 2025 and 2026 research
Project record
No information is collected on this page.
- Permanent project ID
- GP-GE-1E3M4R9
- Catalogued
- 21 Aug 2026
- Completed
- 27 Aug 2026
- Verified
- 27 Aug 2026
- Demonstration
- Included in repository
Handover
After purchase
- 01Payment is confirmed
The project is marked unavailable and cannot be purchased again.
- 02Repository access is granted
The buyer's submitted GitHub account receives access to the private repository.
- 03The purchase record is delivered
The certification sheet is prepared from the reviewed buyer details and sent privately by email.