<p>This study aims to assess the sensitivity of landslide susceptibility mapping (LSM) to various sampling strategies used for non-landslide samples. The study area is Tianshui city, Gansu province, China. Three types of landslide samples, combined with four machine learning models, resulted in a total of 12 scenarios. The receiver operating characteristic curve (ROC), landslide susceptibility index and the mapping distribution characteristics were calculated to access the influences of different sampling strategies and models. The results indicate that the low susceptibility areas sampling strategy yields the highest accuracy for the landslide susceptibility prediction model, followed by the stratified sampling from engineering geological petrofabric (EGP) strategy, and lastly, the random sampling strategy. Analyzing from the perspective of factor importance and the distribution law of landslide susceptibility index under each model, the models employing the stratified sampling from EGP strategy demonstrate greater robustness. In contrast, the models using the random sampling strategy exhibit lower precision and more randomness. In general, the coupled model exhibits strong performance, the frequency ratio coupled adaptive boosting model (FR-AB) demonstrates high sensitivity, while the other models are characterized by their generalizability and robustness. The results reveal the effects of non-landslide sampling strategies and different coupled models on the prediction performance of landslide susceptibility mapping, which provides a reference for subsequent researchers to obtain more reasonable landslide susceptibility mapping.</p>