Optimizing Landslide Susceptibility Mapping with Non-Landslide Sampling Strategy and Spatio-Temporal Fusion Models

Landslides are among the most destructive geological hazards, necessitating precise landslide susceptibility mapping (LSM) for effective risk management. This study focuses on the northeastern region of Leshan City and investigates the influence of various non-landslide sampling strategies and machine learning (ML) models on LSM performance. Ten landslide conditioning factors, selected by SHAP analysis, and six models were utilized: Convolutional neural networks (CNNs), Long Short-Term Memory (LSTM), CNN-LSTM, CNN-LSTM with an attention mechanism (AM), Random Forest (RF), and eXtreme Gradient Boosting combined with Logistic Regression (XGBoost-LR). Three non-landslide sampling strategies were designed, with the certainty factor-based approach demonstrating superior performance by effectively capturing geological and physical characteristics, applying spatial constraints to exclude high-risk zones, and achieving improved mean squared error (MSE) and area under the curve (AUC) values. The results reveal that traditional ML models struggle with complex nonlinear relationships and imbalanced datasets, often leading to high false positive rates. In contrast, deep learning (DL) models—particularly CNN-LSTM-AM—achieved the best performance, with an AUC of 0.9044 and enhanced balance in accuracy, precision, recall, and Kappa. These improvements are attributed to the model’s ability to extract static spatial features (via CNNs), capture dynamic temporal patterns (via LSTM), and emphasize key features through the attention mechanism. This integrated architecture enhances the capacity to process heterogeneous data and extract landslide-relevant features. Overall, optimizing non-landslide sampling strategies, incorporating comprehensive geophysical information, enforcing spatial constraints, and enhancing feature extraction capabilities are essential for improving the accuracy and reliability of LSM.