Efficient probabilistic parameter calibration of landslide run-out models via Bayesian active learning

Landslides, such as debris flows and avalanches, are common natural hazards worldwide. They pose an ongoing threat to life and property. Landslide run-out models that have been developed over the past decades are powerful tools to assess landslide risks and design mitigation strategies. Due to the simplification of real-world landslide processes, the models often contain parameters that rely on calibration of past landslide events where field data are available. Deterministic calibration methods like traditional trial-and-error calibration suffer from the non-uniqueness issue and cannot account for uncertainties associated with field data. Probabilistic calibration methods like Bayesian inference avoid the two issues. However, their usage is hindered by high computational costs due to the long run time of a single run-out model evaluation and the large number of required model evaluations. 

To address the research gap, this work proposes an efficient probabilistic calibration method for parameter estimation of landslide run-out models. The new method couples landslide run-out modeling, Bayesian inference, Gaussian process emulation, and active learning. We implement it in a Python-based environment. Its feasibility and efficiency are tested based on an extensive synthetic case study. Owing to Gaussian process emulation and active learning, our new method overcomes the computational bottleneck by reducing the number of required model evaluations from thousands to a few hundreds. It is therefore expected to advance the state-of-the-art in parameter estimation of landslide run-out models. In addition, the impact of different types of field data on calibration results is studied using the proposed method.