Evaluating the uncertainty in landslide susceptibility prediction: effect of spatial data variability and evaluation unit choices

Traditional landslide susceptibility mapping (LSM) typically employs a point sampling approach, which may neglect the variability of spatial data and the selection of evaluation units, thereby introducing uncertainty into landslide susceptibility predictions. Specifically, when compared to the actual boundary shapes of landslides, simple spatial locations are inadequate for capturing the full spectrum of complex information present in the geological environment, and the correlation between grid units and real-world terrain conditions is not sufficiently close. Addressing these issues, this study focuses on Yongji County as a case study and the spatial coordinates and morphological boundaries of landslides served as input variables for the spatial data, with CatBoost (CB) and Random Forest (RF) algorithms employed for training the predictive models. Subsequently, grid units, slope units and topographic units were selected as mapping units. Ultimately, this study employs analytical techniques such as the Receiver Operating Characteristic (ROC) curve and the analysis of Landslide Susceptibility Indexes (LSI) distributions to assess detailed quantification of uncertainty and precision that results from the selection of spatial datasets and evaluation units. The results indicate that utilizing landslide boundary shapes with higher reliability and precision as input variables significantly enhances the overall accuracy of LSM predictions compared to those based on spatial positions, concurrently diminishing the uncertainty associated with the predictive outcomes; across diverse scenarios, the model that combines slope units with landslide boundary shapes achieves the highest precision.