Landslide displacement prediction using kinematics-based random forests method: A case study in Jinping Reservoir Area, China

Purely empirical and numerical methods are widely used in landslide movement prediction because they can forecast the failure time and consider influence factors, respectively. However, the combination of these two methods for prediction is rare. This paper develops an integrated landslide movement prediction model that can fully consider landslide kinematics and external influence factors using the Verhulst inverse function (VIF) and the random forest (RF) algorithm. The VIF is applied to describe the kinematic behavior of landslides using the rationale of three-stage creep deformation. The RF algorithm is to quantify the response of landslide displacement to the influence of external factors such as reservoir water level and rainfall intensities. The novelty of the VIF-RF model is illustrated by applying to a reservoir landslide, Gapa Landslide, in Southwestern China. The results show that the VIF-RF model shows significant improvement in predicting landslide movement compared with the VIF or RF model. The error analysis confirms that the root mean square error of the VIF-RF decreases by more than 20% compared with the VIF and RF models. In addition, the mean absolute percentage error of the VIF-RF models is less than 5%, a decrease by 2.3% and 10.1% compared to the VIF and RF models, respectively. The feasibility of the VIF-RF model for predicting movement of other reservoir landslides was successfully verified by the Majiagou landslide in the TGRA. The developed VIF-RF model indicates the Gapa landslide deformation is at the primary stage over the monitoring period. The displacement at the G1 and G2 monitoring locations of the Gapa landslide is projected to increase to 2719.8 and 2438.8 mm in January 2021, respectively, and the average rate in the accelerated deformation periods is projected to be 67.8 mm/month. The presented VIF-RF model provides an effective approach for predicting the long-term landslide deformation and identifying its deformation stage.

Automated summary

Empirical and numerical methods are commonly used for landslide movement prediction due to their ability to forecast failure time and consider influential factors. This study introduces an integrated prediction model using the Verhulst inverse function and random forest algorithm to fully consider landslide kinematics and external factors. Results demonstrate significant improvement in prediction accuracy compared to individual models, with a decrease in error rates and increased feasibility for predicting movement of other landslides.