Influence of non-landslide sampling extent on machine learning-based landslide susceptibility assessment, a case study of Wuyi County, China

Non-landslide sample selection plays a critical role in Random Forest-based landslide susceptibility modeling, yet the spatial extent of non-landslide sampling has rarely been systematically quantified. Using Wuyi County (Zhejiang Province, China) as a case study, this study developed 22 non-landslide sampling scenarios with varying spatial extents based on frequency ratio-derived susceptibility zoning, and evaluated their performance using a Random Forest model. Results show that model performance, as measured by the area under the curve (AUC), improves as the non-landslide sampling extent decreases, whereas the lowest performance is consistently observed when the sampling zone approaches full coverage of the study area. By integrating model discrimination metrics with spatial pattern evaluation of susceptibility maps, an optimal non-landslide sampling proportion of 71.42%-89.92% is identified, achieving a balance between predictive accuracy and spatial realism. These findings highlight the sensitivity of susceptibility assessment to non-landslide sampling design and provide a quantitative basis for optimizing sampling strategies in regional landslide susceptibility assessment.