Integrating geospatial data and hybrid machine learning models for landslide susceptibility assessment in semi-arid environments

Landslides are destructive natural hazards with severe impacts on human safety, infrastructure, and the environment. Accurate mapping and prediction of landslide-prone areas are essential for effective risk management and mitigation. This study focuses on the Oued Lakhdar watershed, where three hybrid models were developed by combining Random Forest with Multivariate Adaptive Regression Splines, Naïve Bayes Tree, and Boosted Regression Tree. An inventory of 294 landslides, along with twelve conditioning factors such as elevation, slope, and land use, was used for model training and evaluation. The RF-MARS model identified 14.13% of the study area as highly susceptible to landslides, compared to 11.97% with RF-BRT and 8.23% with RF-NBT. In terms of predictive accuracy, RF-MARS achieved the highest AUC (93.66%), followed by RF-BRT (91.96%) and RF-NBT (91.10%), confirming the superior performance of RF-MARS. These findings demonstrate the value of hybrid machine learning approaches in improving landslide susceptibility assessment and supporting natural hazard management.