Prediction of shear strength parameters of granite residual soil based on Stacking ensemble learning strategy

Granite residual soil is widely distributed in humid and hot regions of southern China, and its highly variable engineering properties bring great challenges to slope stability evaluation and foundation design. The shear strength indices, including cohesion and internal friction angle, are the most critical mechanical parameters for analyzing the stability of geotechnical structures. Traditional laboratory tests for obtaining shear strength parameters are time-consuming, costly and labor-intensive, and cannot meet the demand for rapid parameter acquisition in disaster early warning. In addition, conventional single machine learning models often suffer from limited generalization performance when dealing with the strong nonlinear relationship between soil physical properties and shear strength. To solve the above practical problems, this study develops an innovative prediction framework based on the Stacking ensemble learning algorithm to realize the high-accuracy prediction of cohesion and internal friction angle, and further reveal the dominant influencing mechanism of physical indices on soil shear strength.

In this study, a comprehensive dataset was compiled from published literature and field geotechnical investigation data, and data screening and normalization preprocessing were conducted to unify data quality and eliminate the interference of dimensional differences. A two-layer Stacking ensemble learning architecture was established. Three typical heterogeneous machine learning models—random forest (RF), support vector machine (SVM), and back propagation neural network (BPNN)—were adopted as base learners, and a 5-fold cross-validation strategy was applied to complete model training and avoid overfitting. Ridge regression was employed as the meta-learner to synthesize the prediction outputs from three base learners. Six common geotechnical indices, namely fines content, void ratio, natural water content, liquid limit, plastic limit, and specific gravity, were set as model inputs, while cohesion and internal friction angle were defined as model outputs. Furthermore, the SHapley Additive exPlanations (SHAP) method was introduced to interpret the black-box model and quantitatively analyze the contribution degree of each input parameter.

The results demonstrated that the determination coefficient (R²) of the proposed Stacking model reached 0.88 for cohesion and 0.90 for internal friction angle on the validation set, with corresponding root mean square error (RMSE) values of 3.60 kPa and 2.46°, respectively. Compared with the best-performing single base learner, the R² values increased by 1% and 4%, respectively. Verified by independent engineering test samples collected from Zixing City, Hunan Province, the absolute prediction deviation of cohesion ranged from 0 kPa to 1.91 kPa, and that of internal friction angle varied from 0° to 0.67°. The ensemble model exhibited obviously better prediction capability and robustness than individual models. SHAP interpretation results indicated that cohesion was mainly controlled by water content, liquid limit, and fines content, whereas fines content, void ratio, and water content served as the primary factors affecting internal friction angle. The variation characteristics of all parameters were well consistent with classical soil mechanics theories.

The study proves that the Stacking ensemble learning strategy can effectively combine the respective strengths of different single machine learning models and overcome their inherent defects. The proposed method greatly improves the prediction accuracy and generalization ability for shear strength parameters of granite residual soil. It provides an efficient, low-cost, and reliable technical solution for rapid parameter determination, and has good application prospects in slope stability assessment and landslide disaster prevention in areas covered by granite residual soil.