The accuracy paradox: comparing high-accuracy LiDAR and topographic DEMs for landslide susceptibility assessment in the Slanské vrchy Mountains, Slovakia

Abstract

The Slanské vrchy Mountains, a volcanic mountain range in eastern Slovakia, is a region critical for engineering-geological research due to the widespread occurrence of diverse forms of slope deformation, conditioned by geological structure, erosion, and gravitational processes, as well as by anthropogenic influences. Given the ongoing research, current attention is focused on applying advanced methods for the registration, investigation, and statistical evaluation of slope deformation structures. In this context, the use of a fifth-generation digital elevation model (DEM 5.0), created using airborne laser scanning (LiDAR), is highly relevant due to its sub-metre accuracy. This model provides an excellent basis for generating parametric maps essential for analysing landslide susceptibility. A critical paradox has emerged: DEM 5.0 captures the terrain after a landslide has occurred, including post-failure features such as steep scarps, which can lead to a misleading interpretation of the original slope conditions. In contrast, older, less precise models, such as DEM 3.5 derived from topographic maps, offer a more suitable representation of the pre-landslide terrain. This is because they better reflect the conditions under which landslides initially formed, avoiding the distortion introduced by post-deformation features. This discrepancy significantly impacts parameters such as slope angle, relief curvature and aspect. This study critically examines this issue by comparing parametric maps derived from DEM 3.5 and DEM 5.0 for the southern Slanské vrchy mountains. While the engineering-geological and land cover maps remain consistent across all three analyses, the DEMs yield differing spatial information for elevation, slope, aspect and curvature. Our findings, validated by receiver operating characteristic (ROC) curves and area under the curve (AUC) values, indicate that analyses using DEM 3.5 consistently yielded a higher predictive accuracy (e.g., AUC of 84.18% for the multivariate model) compared to DEM 5.0 (AUC of 80.89%). This highlights that superior accuracy does not always translate to better suitability for specific geomorphological applications. LiDAR (DEM 5.0) excels at mapping and delineating existing slope deformation structures within GIS, but for generating susceptibility maps, it necessitates removing landslide-affected areas and interpolating the original terrain.