Probabilistic Modeling of Landslide Dam Breach Based on Bayesian Inference

The simplified physically-based dam breach models are effective tools for predicting outburst flood hydrographs of landslide dams. However, their predictive reliability is severely constrained by parameter uncertainties, particularly regarding in soil erosion. To address this, this study applies an established Bayesian multilevel framework to develop a probabilistic modeling approach for landslide dam breaches. A highly computationally efficient simplified model is developed and subsequently embedded into a Bayesian multilevel framework to systematically quantify the uncertainties in the erosion parameters. Using observational data from ten documented landslide dam failure cases, model inversion is executed via a Markov chain Monte Carlo (MCMC) simulation combining Gibbs and Metropolis-Hastings sampling. As a primary contribution, this study quantifies the uncertainty of the erosion parameter specifically for landslide dams for the first time. Following inversion, parameters with non-informative priors are updated to well-defined posterior distributions with distinct peaks. Furthermore, the results reveal that approximately two-thirds of the uncertainty in the predicted peak discharge stems from the epistemic uncertainty of key parameters, with the remainder attributed to residual error. This framework significantly improves the reliability of outburst flood predictions and substitutes subjective empirical assumptions with data-driven probabilistic inference, providing highly valuable insights for downstream hazard mitigation.