Bayesian active learning for parameter calibration of landslide run-out models

Abstract

Landslide run-out modeling is a powerful model-based decision support tool for landslide hazard assessment and mitigation. Most landslide run-out models contain parameters that cannot be directly measured but rely on back-analysis of past landslide events. As field data on past landslide events come with a certain measurement error, the community developed probabilistic calibration techniques. However, probabilistic parameter calibration of landslide run-out models is often hindered by high computational costs resulting from the long run time of a single simulation and the large number of required model runs. To address this computational challenge, this work proposes an efficient probabilistic parameter calibration method by integrating landslide run-out modeling, Bayesian inference, Gaussian process emulation, and active learning. Here, we present an extensive synthetic case study. The results show that our new method can reduce the number of necessary simulation runs from thousands to a few hundreds owing to Gaussian process emulation and active learning. It is therefore expected to advance the current practice of parameter calibration of landslide run-out models.

Computationally-feasible uncertainty quantification in model-based landslide risk assessment

“…The runtime of the landslide run-out model is one of the key drivers of computational cost in classical MCS, as it scales with the number of forward evaluations. Gaussian process emulation has been used in recent years to build cheap-to-evaluate emulators to replace expensiveto-evaluate computational models in the framework of uncertainty quantification, such as Sun X. P. et al (2021b), Zeng et al (2021), and Zhao and Kowalski (2022). A Gaussian process emulator is a statistical approximation of a simulation model, built based on input and output data of a small number of simulation runs.…”

“…Such a chain consists of many links, including a digital representation of the topography, the underlying physics-based process model, a numerical solution scheme, the approach to parameter calibration along with the training data it relies on, and concepts used for sensitivity analyses and uncertainty quantification. Challenges in the technical realisation of such integrated workflows have been successfully addressed in the past (Dalbey et al, 2008;Aaron et al, 2019;Sun X. P. et al, 2021b;Zhao et al, 2021;Aaron et al, 2022;Zhao and Kowalski, 2022). It will be of crucial importance in the future to increase the efficiency, sustainability and, hence, acceptance of such orchestrated workflows for landslide risk assessment by improving their robustness, reliability and computational-feasibility.…”

“…They assumed the two friction parameters (Coulomb friction and turbulent friction) in a depthaveraged Voellmy-Salm type approach to follow independent Gaussian distributions, and then studied the induced uncertainty in deposit heights. The major challenge of using MCS for uncertainty quantification in landslide run-out modelling is the high computational cost, as pointed out by many researchers (Dalbey et al, 2008;McDougall, 2017;Aaron et al, 2022;Zhao and Kowalski, 2022).…”

“…The surrogate model can be sampled, instead of the sampling from the forward model, and hence the PDF of the quantity of interest can be calculated efficiently. An example of physics-based machine learning techniques proven to be effective in many applications related to geohazards is Gaussian Process Emulation, with successful demonstrations in landslide run-out models (Zhao et al, 2021;Zhao and Kowalski, 2022) and stability of infrastructure slopes (Svalova et al, 2021).…”

A novel approach for quantifying similarities between different debris flow sites using field investigations and numerical modelling

“…Further, the spatial extent and the impact of flow can be quantified using numerical modelling tools (Mergili et al, 2017; Trujillo‐Vela et al, 2022). However, the applicability of such models is limited in the prediction of future events due to the limitations in calibrating the complex rheological parameters (Zhao & Kowalski, 2022). It is not well studied if the rheological parameters calibrated for one site can be used for the prediction of the shape of debris flows at other sites.…”