Mapping landslide susceptibility using data-driven methods.

Area of Science:

• Geosciences
• Geomorphology
• Environmental Risk Assessment

Background:

• Epistemic uncertainty in landslide susceptibility assessment stems from data errors, causal factor identification challenges, and modeling choices.
• Evaluating these uncertainties is crucial for reliable landslide risk management and spatial planning.

Purpose of the Study:

• To assess the impact of different statistical methods, terrain mapping units, and landslide representation (point vs. polygon) on landslide susceptibility maps.
• To identify optimal modeling strategies for accurate landslide susceptibility assessment.

Main Methods:

• Comparison of logistic regression, discriminant analysis, and information value methods.
• Evaluation of slope, geo-hydrological, and census terrain units.
• Analysis of landslide representation using polygons versus single points.

Main Results:

• Multivariate statistical methods perform best on heterogeneous terrain units.
• Terrain mapping units significantly influence susceptibility results, sometimes more than the statistical method.
• Grid cell units are recommended for accurate landslide inventories, and single points are effective for small landslides.

Conclusions:

• The choice of terrain mapping unit is critical and can introduce greater variability than the statistical method.
• Grid cell terrain units and single-point landslide representation enhance susceptibility map accuracy.
• The Area Under the Curve (AUC) of Receiver Operating Characteristic (ROC) curves may not solely indicate the best model performance, especially with heterogeneous terrain units.