An AutoML and GeoAI workflow for comparative landslide susceptibility assessment in high-relief mountain terrain

Landslide susceptibility mapping in complex mountain terrain demands modeling frameworks capable of resolving non-linear, high-dimensional conditioning factor relationships while maintaining spatial referencing integrity. This study applies and comparatively evaluates the AutoML and GeoAI tools implemented in ArcGIS Pro 3.5 for comparative LSM in Choman District, Kurdistan Region using LightGBM, XGBoost, CatBoost, Random Trees, Extremely Randomized Trees, and a weighted Ensemble model. A spatially balanced training dataset of 2,384 samples was intersected with thirteen conditioning factors and partitioned into 75% training and 25% independent test subsets. On an independent holdout set the Ensemble model returned the lowest log-loss (0.0809) together with an accuracy of 97.15% (95% CI 95.48–98.21%), ROC-AUC of 0.9965, and KS of 0.943; its margin over the leading individual learners was not statistically significant, though it exceeded both bagging-based ensembles on every metric. Under five-fold spatial block cross-validation, adopted as the primary estimate of generalization performance, mean ROC-AUC was 0.993 ± 0.003 and mean accuracy 95.56 ± 1.20%. The three gradient boosting algorithms formed a coherent high-performance cluster, while Tree algorithms showed substantially higher log-loss and compressed probability distributions. Feature importance analysis identified NDVI, land cover, LS factor, and lithostratigraphic units as the four dominant predictors. Hot-spot stability analysis, perturbing the susceptibility index at 10%, 20%, and 30%, showed that 96.84%, 93.70%, and 88.90% of 1,373,610 terrain features respectively retained their original cluster classification, with interquartile ranges of 0 at the two lower magnitudes, indicating that the mapped clustering structure is robust to imprecision in the susceptibility values. This workflow provides a reproducible and uncertainty-aware methodology for similar mountain geohazard situations.