Probabilistic-based prediction of rainfall-induced landslides

Abstract

Prediction of rainfall-induced landslides has received considerable attention amongst the scientific community due to the geological hazard’s catastrophic impacts. The prediction is commonly performed based on rainfall threshold. However, less attention has been given to physical-based thresholds. The thresholds are also mainly determined based on deterministic model. The inherent uncertainties in soil properties are neglected. Therefore, this study aims to improve the prediction of landslides in unsaturated slopes by incorporating the uncertainties in soil properties. The performance of the landslide predictive models can be enhanced towards a more reliable landslides warning system. One of the major slope failure events in Kota Kinabalu, Sabah, Malaysia, is selected as a case study. Statistical analyses have been conducted to characterize the uncertainties in hydro-mechanical soil variables by identifying best-fitted marginal distribution amongst normal, lognormal, Gumbel, and Weibull distribution. The dependencies of the multivariate are assessed using different types of vine copula models. Then, a reliability-based probabilistic analysis has been proposed to determine the performance level of the slope by integrating the Monte Carlo Simulation and Multilayer Perceptron regressor, using 120 samples of soil properties generated from the Latin Hypercube Sampling. Three types of rainfall thresholds, namely intensity-duration, cumulative rainfall-duration, and daily rainfall-antecedent rainfall for various antecedent days of 5, 10, 15, 20, 25, and 30 days are proposed. Comparison of rainfall threshold based on probabilistic and deterministic models shows that the former outperforms the latter in threat score. The antecedent rainfall of 10 and 15 days can well describe the landslides initiation compared to other antecedent rainfall durations for the daily rainfall-antecedent rainfall threshold. This study mainly contributes to the development of a new physical-based rainfall threshold for predicting landslides initiation using a reliability-based probabilistic approach by incorporating the uncertainties in dependent hydro-mechanical soil variables for the first time.