Evaluating scale effects of topographic variables in landslide susceptibility models using GIS-based machine learning...

Area of Science:

• Geosciences
• Geomorphology
• Remote Sensing

Background:

• Digital Elevation Models (DEMs) are crucial for geomorphic studies, but their quality and resolution effects on landslide susceptibility mapping (LSM) are not well understood.
• Assessing the scale dependency of DEM-derived factors is vital for accurate hazard assessment.

Purpose of the Study:

• To investigate the influence of DEM resolution and quality on landslide susceptibility mapping.
• To determine the optimal DEM resolution for generating geomorphometric factors for LSM.

Main Methods:

• Utilized a 5m LiDAR DEM, a resampled 30m LiDAR DEM, and a 30m ASTER DEM for geomorphometric factor derivation.
• Compiled a landslide inventory for Sihjhong watershed, Taiwan (2004-2014) with 267 events.
• Applied logistic regression (LR), random forest (RF), and support vector machine (SVM) for LSM, evaluating accuracy using overall accuracy, kappa index, and ROC curves.

Main Results:

• The resampled 30m LiDAR DEM derivatives produced the highest accuracy in LSM, contrary to the expectation that finer resolution is always better.
• Random Forest (RF) demonstrated superior performance compared to logistic regression (LR) and support vector machine (SVM) across the tested DEM resolutions.
• Scale dependency of DEM-derived factors significantly influences LSM outcomes.

Conclusions:

• The choice of DEM resolution is a critical factor in landslide susceptibility mapping, and higher resolution does not always guarantee better results.
• Random Forest is a robust machine learning model for LSM.
• Findings provide guidance for selecting appropriate DEM resolutions for landslide hazard assessment in vulnerable regions.