Decadal landslide susceptibility mapping: Impacts of sampling methods on prediction accuracy

Landslide susceptibility mapping (LSM) is crucial for reducing disaster risk in complex mountainous regions. This study evaluated the impact of various sampling methods on the accuracy of LSM over the next decade in Bijie City, Guizhou Province, China. Datasets were collected from 614 landslides and 500 non-landslides, and four sampling methods were proposed. Recurrent Neural Network (RNN), Gated Recurrent Unit (GRU), K-Nearest Neighbor (KNN), and Extreme Gradient Boosting (XGB) models were assessed utilising 15 metrics (Elevation, Slope, Aspect, Plan curvature, Profile curvature, Stream Power Index, Sediment Transport Index, Vector Ruggedness Measurement, Topographic Roughness Index, Lithology, Land use, Normalized Difference Vegetation Index (NDVI), Rainfall, Distance from Road, Distance from River). The results demonstrated that the GRU model combined with a 5-m sample boundary from the interior of the landslide and non-landslide areas exhibited superior performance with F1, Accuracy, and Area Under Curve (AUC) scores of 0.9700, 0.9450, and 0.8925, respectively. LSM will be projected for the next decade by coupling the Geophysical Fluid Dynamics Laboratory Earth System Model version 4 (GFDL-ESM4) with the Shared Socioeconomic Pathway (SSP119). This study provides valuable insights into landslide risk management in landslide-prone areas.