X-SEL: a stacked ensemble learning architecture for forecasting soil slide displacement in flume simulations

Abstract Background Predicting soil sliding displacement plays an essential role in developing landslide early-warning systems. Even though physics-based approaches, e.g., trapezoidal numerical integration, have demonstrated promising predictive performance under ideal laboratory conditions, they are very susceptible to field disturbances such as sensor noise and drift. They fail to adequately model the nonlinear behavior and transition phenomena of soil sliding events. Machine learning algorithms (ML), which are promising solutions for predicting soil sliding displacement due to their ability to handle nonlinearity, suffer from dependence on dataset properties. Sample size, noise levels, and feature correlations play major roles in determining model accuracy. This research presents an XGBoost-based stacked ensemble learning algorithm called X-SEL for predicting soil sliding displacement using IMU sensor data from flume experiments. Results This study investigates the capability of an XGBoost-based ensemble learner for soil sliding displacement predictions using three experimental datasets. Three different datasets were used for validating the X-SEL framework that correspond to varying sliding regimes, sample sizes, and feature correlations. The X-SEL framework had the lowest average root mean square error (RMSE), equal to 0.2714, compared with the benchmark models among the three datasets examined. Nevertheless, other individual learners, including kNN and XGBoost produced competitive results. Hence, it can be concluded that X-SEL does not always have higher predictive power than each of its base-learners. Nevertheless, X-SEL proved superior compared with the traditional trapezoidal integration method in replicating smoothly transitioning non-linear displacements. Furthermore, SHAP analysis suggested that the meta-learner adaptively weights reliable base-learner models, and triaxial acceleration features contribute more to displacement prediction than gyroscopic features. Conclusion The X-SEL framework developed in this paper serves as a proof-of-principle of utilizing inexpensive IMU sensors for predicting soil sliding displacement. Future studies will aim at expanding X-SEL beyond the flume environment, conducting uncertainty quantification, and designing practical tools for deployment purposes.