Reducing subjectivity in landslide susceptibility mapping: machine learning-informed multi-criteria analysis with spatial uncertainty assessment

Landslide susceptibility mapping faces the challenge of subjective weight assignment in Multi-Criteria Decision-Making (MCDM) approaches, while machine learning methods lack transparency in factor importance. This study presents a hybrid approach integrating Forest-Based Classification and Regression (FBCR) with MCDM techniques in Qasri District, Kurdistan, Iraq, a transitional zone between Zagros Mountains and foothills with complex topography. Using 883 mapped landslides and 15 conditioning factors from multi-source geospatial data, we developed three susceptibility models: FBCR, SAW, and TOPSIS. FBCR achieved strong performance (R 2 = 0.93 training, 0.92 validation) while generating objective variable importance scores used as MCDM weights. The LS-factor was the most influential predictor (14%), followed by NDVI (11%) and land cover (10%). Validation using ROC analysis showed SAW achieving highest performance (AUC = 0.933), followed by FBCR (0.908) and TOPSIS (0.882). However, spatial distributions differed markedly: FBCR classified 64% of the area as very high susceptibility, SAW 37%, and TOPSIS 15%, reflecting their contrasting aggregation logics. Uncertainty analysis using the Assess Sensitivity to Attribute Uncertainty with a triangular simulation and uniform ±5/10/20% perturbations (100 iterations) revealed stability decline from 92.02% (±5%) to 83.01% (±10%) and 70.41% (±20%), while IQR and SD of outcomes widened, pinpointing stable core hotspots whereas marginal hillslopes were most sensitive. Overall, FBCR produced a high-confidence susceptibility map with interpretable drivers while SAW yielded the sharpest class stratification from the same factors and weights. Crucially, embedding simulation-based uncertainty within the same GIS environment transformed a deterministic map into a confidence-aware product suitable for planning and risk management.