Comparison of different machine learning models coupling with logistic regression for landslide susceptibility mapping

Landslides are among the most destructive geological hazards, highlighting the urgent need for accurate landslide susceptibility mapping (LSM) to support risk reduction and mitigation strategies. Here, we systematically assess the performance of individual machine learning (ML) models and their logistic regression (LR)-coupled counterparts, with a particular focus on the influence of raster resolution on model accuracy. A total of 10 landslide conditioning factors were selected to construct both individual and coupling models, while correlation analysis and SHAP-based feature attribution were applied to ensure input independence and enhance interpretability. Hyperparameters were optimized via Bayesian search. Results indicate that slope, lithology, and elevation exert the strongest controls on landslide occurrence, and that deep learning (DL) architectures outperform other individual models. Crucially, all LR-coupled models yielded significant gains over their standalone equivalents, with AUC improvements of 4.4% (DNN_LR), 6.1% (BP_NN_LR), 6.0% (XGBoost_LR), 3.9% (RF_LR), and 5.1% (SVM_LR). DL-based hybrids achieved the highest predictive accuracy, although LR tended to overpredict low-risk zones. Across multiple raster resolutions, coupled models, particularly DNN_LR and BP_NN_LR, exhibited strong robustness and spatial generalizability. Overall, we propose a novel LR-ML coupling framework that integrates the transparency and efficiency of LR, a lightweight model with superior linear modeling capacity, with the representational power of non-linear meta-learners (RF, SVM, XGBoost, BP_NN, and DNN). LR provides efficient preliminary predictions and refines label quality via targeted non-landslide sampling, yielding high-quality training inputs for subsequent learning. This integration effectively mitigates overfitting, enhances interpretability, and reduces computational demand, while maintaining stability across scales. Collectively, our findings establish LR-ML as a robust and scalable framework for large-scale LSM.