Synergistic evolution of hydrological and movement characteristics of Majiagou landslide and identification of key triggering factors through interpretable machine learning

Abstract

Variations in reservoir water level and seasonal precipitation have reactivated or accelerated numerous reservoir landslides in the Three Gorges Reservoir (TGR) area in China since its impoundment in 2003. Majiagou landslide, a typical reservoir landslide with stabilizing piles, is affected by the coupling effect of rainfall and reservoir level fluctuations. Monitoring data of nearly 11 years show continuous movement of Majiagou landslide, in contrast to the step-like movements of many landslides in this region. Displacements of the landslide surface and sliding zone are accelerated in rainy seasons accompanied by rapid fluctuations in reservoir water level. A SHAP-XGBoost-based interpretable machine learning method was proposed to identify the key triggering factors of the deformation of Majiagou landslide. The crucial triggering factors vary among different monitoring sites, monitoring periods (e.g., before and after the replacement of monitoring sites), and monitoring intervals. Rainfall makes the most prominent contribution to the displacements of the landslide surface and slip zone. From the front to the rear of Majiagou landslide, the response period of surface deformation to reservoir water level fluctuation gradually lengthens, and the middle and rear parts are more sensitive to the average reservoir water level in the short term. The proposed SHAP-XGBoost method will facilitate deformation prediction, stability evaluation, and the calibration of early warning systems for reservoir landslides.