Influence of sample purity on landslide susceptibility mapping: a mutually exclusive stratified purity strategy based on the state-process consistency

The quality and uncertainty of training samples are a critical bottleneck constraining the performance of machine learning in landslide susceptibility mapping (LSM). To address this, this study introduces a sample purity optimization framework based on state-process consistency (SPC). The framework utilizes two indices, SWCFD and CF-BE-HMD, combined with a mutually exclusive stratified purity strategy (MESPS), to systematically optimize the purity of the training samples. The framework was validated using data from Zhenxiong County, Yunnan Province, China. The results indicate that sample purity is the core variable influencing LSM model performance. The mutually exclusive sample group with the highest purity achieved AUC values of 0.92–0.93, markedly higher than those of other groups. High-purity landslide and non-landslide samples effectively suppressed the risk generalization effect caused by low-purity positive samples and overcame the risk conservatism problem induced by low-purity negative samples, resulting in clearer boundaries in the risk zonation. Therefore, only through bidirectional purification, which ensures that both positive and negative samples are highly representative, can these two extreme biases be maximally eliminated, leading to the generation of scientific, reliable, and balanced landslide susceptibility maps. This study offers a feasible solution to the sample quality bottleneck, providing a solid theoretical and empirical basis for shifting landslide modeling towards a quality-driven approach.