Inventory Incompleteness-Induced Uncertainty in Machine Learning-Based Landslide Susceptibility Modeling: A Comparative and Interpretable Analysis

Landslide susceptibility prediction is widely used for regional hazard assessment, but its reliability depends on the completeness of landslide inventory data. This study examined how inventory loss affected model performance, spatial predictions, and fitted-model attribution in Fuyang District, Hangzhou, China. Four nested random-removal conditions (10%, 20%, 30%, and 40%) were evaluated using Random Forest (RF), Multilayer Perceptron (MLP), and Logistic Regression (LR). The analysis combined AUC, Landslide Susceptibility Index statistics, spatial differences, susceptibility-class area proportions, ten repeated runs, and SHAP attribution. Frequency Ratio (FR) values were recalculated for each inventory condition before the train–validation split and five-fold cross-validation. The repeated runs showed model-specific, non-monotonic AUC responses. Mean AUC values across the missingness conditions were 0.732–0.749 for RF, 0.666–0.779 for MLP, and 0.732–0.772 for LR. Two additional RF tests represented accessibility-related omission and spatially clustered omission. Their AUC values were 0.723 and 0.672, respectively. A control analysis with unchanged raw predictors gave results close to the FR input at 10% and 20% missingness, with larger differences at 30% and 40%. Spatial differences also showed that a higher AUC did not always mean closer agreement with the model-specific assumed ideal map. SHAP results showed changes in fitted-model feature attribution, not changes in the physical landslide-forming process. The findings are limited to this regional dataset, the tested omission rules, and the current validation design.