Hydrometeorological Control Sampling for Rainfall Induced Landslide Susceptibility Modelling Using Multi Source Data and Advanced Learning

Rainfall-induced landslides result from interactions between terrain predisposition and hydrometeorological forcing. In event-scale susceptibility modelling, uncertainty often arises from non-landslide controls. Locations without recorded failures may differ in rainfall history, storm exposure, or inventory completeness, which can bias models trained on static absence samples. This study develops a hydrometeorological control sampling strategy for rainfall induced landslide susceptibility modelling. The strategy defines each sample by grid cell and rainfall date, and constructs non landslide controls from storm related risk sets. A background predisposition prior is used to screen candidate controls. Regional same date controls, annular hard controls near failed slopes, and cross year rainy season background controls are then integrated to represent complementary hydrometeorological and terrain conditions. Design weights and density ratio calibration are applied to account for control reliability and reduce distribution mismatch between the training sample and the mapping domain. In the sample-level evaluation, the method was evaluated in Pubei County, Guangxi, China, using terrain, geology, land cover, daily and antecedent rainfall, and surface wetness. It outperformed Buffer, LowSlope, and IV Low across five classifiers. Relative to IV Low, mean AUC increased from 0.887 to 0.958, accuracy from 81.8% to 91.6%, and Kappa from 63.6% to 83.3%. Holdout validation of two July 2006 landslide clusters also showed greater concentration in top-ranked areas. Averaged over RF and GBDT, top 10% capture rose from 0.227 to 0.322, while the frequency ratio increased from 2.264 to 3.213. These findings suggest that, under the evaluated conditions, the strategy improves sample discrimination, increases landslide concentration in areas ranked as highly susceptible, and reduces uncertainty in the selection of nonlandslide controls.