Dual tree-boosting framework for estimating warning levels of rainfall-induced landslides

Abstract

The early warning system (EWS) has become an efficient approach to mitigating rainfall-induced landslide risk. An accurate estimation of landslide probability given rainfall events majorly determines the system's success. Recent studies show significant effects of soil hydrological information on the forecast quality of landslide occurrence. However, a common challenge is a lack of pertinent information on in situ hydrological conditions. This study develops a dual tree-boosting framework capable of addressing near real-time pore water pressure while predicting warning levels of rainfall-induced landslides with minimum input requirements. The framework was implemented by compiling the extreme gradient boosting (XGB) model for estimating pore water pressure and the categorical boosting (CatBoost) model for classifying warning levels. Bayesian reasoning is coupled with the K-fold cross-validation to optimize these models' hyperparameters. The presented framework was applied to the case study of landslides in the Boso Peninsula, Japan, with contrasting destabilization mechanisms to demonstrate the feasibility of operating the local scale EWS. The XGB-predictive model showed excellent performance by satisfying the consistency criterion with a uniform and narrow prediction interval. Also, the reliability of warning signals for the critical conditions for landsliding under different geological conditions issued by the proposed framework was verified by the reasonably low error rate (i.e., ER = 0.24-1.31%). Given the high efficiency with simplified inputs and self-improvability, the presented framework demonstrates excellent applicability in the operational EWS in response to the increasingly pronounced climate change impacts.