Objectives Landslide susceptibility mapping (LSM) serves as a fundamental tool for geohazard prevention and regional planning. However, the accuracy of LSM in complex geological environments, such as Baoji City, is often constrained by two primary challenges: The scarcity of historical landslide samples leading to poor spatial representation, and the limitation of traditional convolutional neural networks (CNN) in capturing long-range spatial dependencies. To address these issues, we propose a novel framework that integrates a positive sample buffering strategy with a hybrid Transformer-CNN model. The primary objective is to enhance the spatial representation of landslide samples and effectively fuse local and global features to improve the predictive performance and interpretability of LSM.

Methods First, to overcome the limitations of point-based sampling which fails to reflect the actual spatial extent of landslides, a dynamic buffering strategy was implemented. Based on the scale of historical landslides, variable buffer radii ranging from 90 m to 130 m were applied to expand positive samples from isolated points into spatial polygons. Second, a total of 13 conditioning factors were selected, including topography (elevation, slope, aspect, plan curvature, profile curvature), geology (lithology, distance to faults), hydrology (distance to rivers, topographic wetness index, rainfall), ecological environment (normalized difference vegetation index) and human activities (land use, distance to roads). Multicollinearity analysis, utilizing both Pearson correlation coefficients and variance inflation factor (VIF), was conducted to ensure the statistical independence of these factors. Third, a Transformer-CNN coupled model was constructed. The architecture utilizes a CNN backbone with a 15×15 input window to extract high-dimensional local features, which are then serialized and fed into a Transformer encoder. The Transformer employs a multi-head self-attention mechanism to dynamically weigh the importance of different spatial positions, thereby capturing global non-linear interactions among factors. To rigorously evaluate the proposed framework, a total of eight experimental scenarios were designed, comparing the performance of random forest, standalone CNN, standalone Transformer, and the coupled Transformer-CNN model under both buffered and unbuffered sample conditions. Furthermore, the Shapley additive explanations (SHAP) algorithm was integrated to interpret the "“black-box”decision process of the optimal Transformer-CNN model, quantifying the contribution of each factor to landslide susceptibility.

Results Experimental validation demonstrate significant improvements in both model robustness and prediction accuracy. (1) Effectiveness of buffering strategy: The application of the buffering strategy universally enhanced the performance of all tested models. Specifically, the area under the curve values for RF, CNN, Transformer, and Transformer-CNN increased from 0.834, 0.852, 0.847, and 0.875 (unbuffered) to 0.883, 0.913, 0.926, and 0.959 (buffered), respectively. (2) Superiority of the hybrid model: The proposed Transformer-CNN model consistently outperformed single-structure models in both experimental settings. In the unbuffered scenario, it improved the AUC from 0.852 (CNN) and 0.847 (Transformer) to 0.875; in the buffered scenario, it further elevated the AUC from 0.913 (CNN) and 0.926 (Transformer) to 0.959. (3) Interpretability analysis: Based on SHAP interpretability analysis, it was further revealed that lithology, annual precipitation, and slope aspect contributed most significantly to predicting landslide susceptibility in Baoji City, with contributions of 0.55, 0.47, and 0.43, respectively. The interaction effects among these three factors were significant.

Conclusions The positive sample buffering strategy effectively mitigates the issue of insufficient spatial representation in historical inventories. The Transformer-CNN coupled architecture successfully overcomes the intrinsic limitations of single-structure models, offering a state-of-the-art solution for complex non-linear geological modeling. This study establishes a robust framework for landslide susceptibility assessment in data-scarce regions.These findings provide actionable scientific insights for local disaster mitigation and land-use planning.