Gaussian Process Emulation for Landslide Run-out Modeling

Abstract:

Shallow flow type landslide run-out models have been greatly developed over the past decades and been successfully utilized for the purpose of hazard assessment and mitigation. However, computational costs—namely relatively long model run times—limit their usage, especially when it comes to computationally intensive tasks like sensitivity analysis, uncertainty quantification [1], and model calibration. The aim of this study is to take advantage of the recent progress of Gaussian process emulation in the machine learning community to address this issue. The novel R package RobustGaSP (Robust Gaussian Stochastic Process Emulation) [2] and the innovative open-source mass flow simulation software r.avaflow [3] are integrated in a unified Python-based framework. In this way a computationally efficient Gaussian process emulator can be conveniently trained based on a modest number of landslide run-out model evaluations. It is then used as a surrogate model for tasks in which directly employing the landslide run-out models is computationally impractical. The methodology is validated and its usage is illustrated based on past landslide events. Further discussions are devoted to properly treating the uncertainty introduced by Gaussian process emulation and employing the machine learning enhanced process models in above-mentioned computationally demanding tasks.

[1] Zhao, H., Kowalski, J., 2020. Topographic uncertainty quantification for flow-like landslide models via stochastic simulations. Natural Hazards and Earth System Sciences 20, 1441–1461. doi:10.5194/nhess-20-1441-2020.

[2] Gu, M., Palomo, J., Berger, J.O., 2019. RobustGaSP: robust Gaussian stochastic process emulation in R. The R Journal 11, 112–136. doi:10.32614/RJ-2019-011.

[3] Mergili, M., Fischer, J.T., Krenn, J., Pudasaini, S.P., 2017. r.avaflow v1, an advanced open-source computational framework for the propagation and interaction of two-phase mass flows. Geoscientific Model Development 10, 553–569. doi:10.5194/gmd-10-553-2017.