Effects of DEM Resolution and Mapping Units and Their Interaction on Landslide Susceptibility Assessment: A Case Study from Shangyou County, Jiangxi, China

DEM resolution and mapping-unit selection influence terrain representation and landslide susceptibility assessment. However, the relative effects and interaction remain insufficiently understood in hilly regions dominated by shallow and small-scale landslides. Shangyou County, Jiangxi Province, China, was selected as the study area. The region is characterized by widespread shallow landslides and intensive slope-cutting activities. Two DEM datasets (5 m and 12.5 m) were combined with two mapping-unit types (grid cells and slope units) to establish four evaluation scenarios. Ten environmental factors, including elevation, slope, aspect, overburden thickness, engineering geological groups, distance to roads, distance to rivers, distance to faults, cut-slope density (CutDensity), and mean annual rainfall were used as model inputs. CutDensity was derived from 11912 mapped cut-slope polygons recorded irrespective of landslide occurrence and was used to represent anthropogenic disturbances. Model performance was assessed using repeated five-fold cross-validation and spatial block cross-validation, ROC-AUC, and feature-importance analysis. Differences between the original DEM products varied with mapping-unit type, with relatively small differences for grid-cell models but larger differences for slope-unit models. A significant DEM-mapping-unit interaction was observed for both RF and XGBoost. Controlled resampling showed generally small AUC differences between the original 5 m DEM and its 12.5 m resampled version, suggesting that differences between original DEM products cannot be attributed to nominal resolution alone. Feature-importance analysis showed that CutDensity ranked first in RF, whereas GeoGroup had the highest importance in XGBoost. Removing CutDensity reduced the mean AUC by 0.0293 for RF and 0.0240 for XGBoost. Map comparisons focused on spatial agreement, class transitions, and observed landslide proportions, recognizing the limitations of quantile-based classification. Within the two algorithms evaluated, changes associated with DEM-mapping-unit combinations were generally larger than differences between RF and XGBoost. The results suggest that DEM source characteristics, mapping-unit design, and spatial dependence should be considered when interpreting susceptibility results.