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

Introduction: Increasing complexity and capacity of computational physics-based landslide run-out modelling yielded highly efficient model-based decision support tools, e.g. landslide susceptibility or run-out maps, or geohazard risk assessments. A reliable, robust and reproducible development of such tools requires a thorough quantification of uncertainties, which are present in every step of computational workflow from input data, such as topography or release zone, to modelling framework used, e.g. numerical error.

Methodology: Well-established methods from reliability analysis such as Point Estimate Method (PEM) or Monte Carlo Simulations (MCS) can be used to investigate the uncertainty of model outputs. While PEM requires less computational resources, it does not capture all the details of the uncertain output. MCS tackles this problem, but creates a computational bottleneck. A comparative study is presented herein by conducting multiple forward simulations of landslide run-out for a synthetic and a real-world test case, which are used to construct Gaussian process emulators as a surrogate model to facilitate high-throughput tasks.

Results: It was demonstrated that PEM and MCS provide similar expectancies, while the variance and skewness differ, in terms of post-processed scalar outputs, such as impact area or a point-wise flow height. Spatial distribution of the flow height was clearly affected by the choice of method used in uncertainty quantification.

Discussion: If only expectancies are to be assessed then one can work with computationally-cheap PEM, yet MCS has to be used when higher order moments are needed. In that case physics-based machine learning techniques, such as Gaussian process emulation, provide strategies to tackle the computational bottleneck. It can be further suggested that computational-feasibility of MCS used in landslide risk assessment can be significantly improved by using surrogate modelling. It should also be noted that the gain in compute time by using Gaussian process emulation critically depends on the computational effort needed to produce the training dataset for emulation by conducting simulations.