A novel approach incorporating credibility and spatial uniformity for landslide negative sampling

Landslide susceptibility assessment (LSA) plays a crucial role in the prevention and control of landslide disasters, and selection of negative-sample directly affects the model prediction performance of models. Previous methods encounter difficulty in simultaneously considering sample credibility and spatial uniformity, which restricts the ability of models to learn non-landslide features and affect the accuracy. To address this issue, this study proposes a negative sampling method based on credibility and spatial-uniformity (CSUN). This method first divides landslide feature spaces using K-means clustering, and uses a Gaussian similarity function to measure the environmental similarity between each clustering unit and landslide samples, constructing a negative sample credibility map and selecting negative samples accordingly. Random Forest (RF) and Support Vector Machine (SVM) models were used to compare CSUN with random and credibility-only sampling methods. The results show that: (1) Compared with random and credibility-only negative-sample selection schemes, the proposed strategy improved the AUC by 20.5% and 8.7% under the RF model, and by 12.3% and 6.5% under the SVM model. (2) Repeated sampling analysis demonstrates that CSUN produced the lowest standard deviations of AUC and F1-score, indicating the most stable performance. The proposed approach improves the prediction accuracy of the models and provides an effective strategy for constructing reliable and spatially uniform landslide negative samples.