Landslide susceptibility mapping through the integration of pixel and slope unit analyses: a case study in El Salvador, Central America

Landslide susceptibility mapping is commonly performed using either pixel-based or slope-unit (SU)-based approaches, each characterised by specific advantages and limitations. In this study, SU-level information is incorporated directly into a pixel-based modelling framework to assess whether this integration improves model performance and the spatial organisation of susceptibility maps. The analysis focuses on the northern sector of the Ilopango Caldera (El Salvador), where shallow landslides were triggered by Tropical Storms Amanda and Cristobal in 2020. A 2 m resolution LiDAR-derived DEM was used to derive pixel-scale terrain predictors and to delineate slope units through hydrological analysis. Two configurations were tested: a Baseline model based solely on pixel-scale variables and an Enhanced model including SU-based morphometric summaries and lithological proportions. Both configurations were implemented using Multivariate Adaptive Regression Splines (MARS) and Random Forest (RF) and evaluated through repeated spatial cross-validation. The Enhanced models systematically outperformed the Baseline models. Mean AUC increased from 0.927 to 0.942 for MARS and from 0.923 to 0.949 for RF. In addition to the improvement in predictive performance, the Enhanced configuration produced maps characterised by higher spatial autocorrelation, measured using Global Moran’s I, and reduced local fragmentation, while maintaining variability within individual slope units. These results indicate that integrating SU-derived descriptors during model calibration strengthens the predictive capability of pixel-based models and improves the spatial consistency of the resulting maps, without sacrificing pixel-scale detail.