Enhancing the performance of medium-resolution DEM-based landslide susceptibility maps with machine learning algorithms: a case study of Watauga County, North Carolina

Abstract

The performance of landslide susceptibility maps relies heavily on the quality and resolution of the digital elevation model (DEM) utilized for deriving the causal factors of landslides. In many regions of the world, the unavailability of high-resolution DEMs necessitates the use of medium resolution (30 m) freely available DEMs. This study evaluates and compares the predictive performance, classification alignment, and spatial agreement of landslide susceptibility maps generated from three 30 m DEMs, ASTER, EDNA, and SRTM with a 1 m LiDAR DEM and introduces two integration approaches, an equal-weight linear combination and a Principal Component Analysis (PCA)-based method, to improve mapping accuracy where high-resolution data are unavailable. Using two machine learning algorithms, Random Forest and Extreme Gradient Boosting (XGBoost), we conducted our analysis in Watauga County, North Carolina. We incorporated seven DEM-derived factors alongside six non-DEM-based factors. We evaluated models based on predictive performance, spatial agreement, and classification alignment. The 1 m DEM–based models produced the highest AUC scores (95.01% for Random Forest and 99.81% for XGBoost), outperforming the individual 30 m DEM models by as much as 2.35%. Notably, the linear weighted combination of the three 30 m DEMs (AUC: 99.10% for XGBoost) and the PCA-based integration approach both enhanced predictive performance, spatial agreement (up to 99.69% for Random Forest and 83.63% for XGBoost), and classification alignment (up to 90.73% for Random Forest and 82.15% for XGBoost) relative to the 1 m DEM based reference. Based on these results, we recommend applying a linear weighted combination of medium-resolution DEMs for landslide susceptibility mapping in regions where high-resolution DEMs are not available.