Abstract: To address the issues of high costs associated with shear strength testing of granite residual soil—a critical medium for landslides—and the limited prediction accuracy of single machine learning models, a shear strength prediction method based on the Stacking ensemble strategy is proposed. This method integrates three heterogeneous base learners: Random Forest (RF), Support Vector Machine (SVM), and Back Propagation Neural Network (BPNN). A two-layer Stacking framework is constructed using five-fold cross-validation, with logistic regression selected as the meta-learner. Based on geotechnical test data, six parameters—fines content, void ratio, water content, liquid limit, plastic limit, and specific gravity—were chosen as input features, while cohesion and internal friction angle served as output parameters. The model's performance was systematically evaluated, and its decision-making mechanism was interpreted using the SHAP method.The results indicate that the Stacking ensemble model achieves R² values of 0.88 and 0.90 for cohesion and internal friction angle on the validation set, with RMSE values of 3.60 kPa and 2.46°, respectively. These represent improvements of 1% and 4% in R² compared to the best single models (BPNN for cohesion and SVM for friction angle). On the test set, the prediction deviation ranges are 0–1.91 kPa for cohesion and 0°–0.67° for the internal friction angle, both outperforming the comparison models. SHAP analysis reveals that the key influencing factors for cohesion are, in order, water content, liquid limit, and fines content; for the internal friction angle, the dominant factors are fines content, void ratio, and water content. These findings are highly consistent with soil mechanics theory.The utilization of the Stacking ensemble learning strategy effectively combines the advantages of multiple models, significantly enhancing the prediction accuracy and generalization capability for the shear strength of granite residual soil. This provides a rapid and low-cost technical approach for landslide hazard prevention and control in areas where granite residual soil is distributed.