Is there difference in landslide susceptibility model based on explainable artificial intelligence from the perspective of slope units with different scales?

The scale of mapping units significantly affects the accuracy and reliability of landslide susceptibility assessment. However, existing landslide susceptibility studies lack a clear determination of the appropriate slope unit scale, and the impact of different slope unit configurations on the modeling process and model interpretability has not been thoroughly investigated. This study conducted an empirical analysis using extensive real-world landslide data from the core area of the Three Gorges Reservoir region, comprehensively investigating the effect of slope-unit scales on the landslide susceptibility assessment. Initially, a geospatial dataset comprising 3594 historical landslide events and 22 initial condition factors was compiled. Subsequently, 30 different slope unit schemes of varying scales were generated by the r.slopeunits tool. For each scheme, the dataset was randomly divided into training and testing subsets with a 7:3 ratio and modeled using random forest model. This study reveals the significant impact of slope unit scales on hyperparameter optimization, factor selection, and model interpretability. The results highlight that: (1) appropriate slope unit scale can improve the quality of input variables, thereby enhancing the generalization ability and interpretability of landslide susceptibility assessments, reducing the risk of overfitting. (2) Finer and more concentrated slope units do not always lead to better results; they may excessively rely on distance metrics, resulting in overly conservative high susceptibility classifications in landslide susceptibility models. This study provides valuable insights into selecting the appropriate slope unit scale for landslide susceptibility assessment.