UC Berkeley

Major earthquakes can trigger hundreds to thousands of landslides in mountainous regions, posing significant threats to infrastructure, property, and human lives. Developing effective regional landslide back-calculation (or inversion) and prediction methods is essential for enhancing community resilience against landslides during seismic events. However, current approaches for analyzing earthquake-induced landslides at the regional scale are underdeveloped. Available prediction approaches use primarily grid cell (or pixel)-based calculations for computational efficiency but lack the ability of deriving individual landslide size characteristics. Other published models include landslide geometry but rely on event statistics of the complete landslide inventory for inversion. These models do not consider the location and geometry of individual landslides and are unable to distinguish the independent contributions of cohesion and friction to the strength of near-surface materials. To address these disadvantages, a back-calculation or inversion approach is developed as a contribution to computationally-efficient regional earthquake-induced landslide prediction analysis.

A pseudo-three-dimensional (pseudo-3D) procedure is used to develop landslide failure surfaces for use with inversion and forward prediction applications. The procedure relies on one-dimensional (1D) seismic displacement models that are used to identify triggering locations from a digital elevation model (DEM) raster, which are then projected into three-dimensional (3D) cone-shaped failure surfaces using an upslope projection of the failure surface. An inversion methodology is proposed to back-analyze earthquake-induced landslide inventories with numerous landslides using the pseudo-3D procedure. The methodology is named “RESTAB-inversion” (REgional Seismic STABility Analysis). The RESTAB-inversion algorithm identifies a best-matching modeled landslide geometry by minimizing the mismatch of landslide location, area, and volume between modeled and mapped landslides. The output includes a 3D best-matching modeled landslide and an estimate of average shear strength with uncertainty for each mapped landslide. Using a collection of multiple landslides of varying sizes (i.e., failures occurring under a range of normal stresses), shear strength estimates can be aggregated to derive regional-scale Mohr-Coulomb strength envelopes using geologic unit information or K-means clustering. The potential bias of the proposed methodology is assessed by inverting synthetic landslide inventories generated using a 3D limit equilibrium method, which produces similar back-calculated friction angles but lower cohesion values than those from the 3D slope stability analysis.

Recent earthquake-triggered landslide events with high-resolution mapping offered the opportunity to employ RESTAB-inversion for regional estimates of strength parameters using K-means clustering. Two earthquake events are used as examples: the 2015 Mw 6.5 Lefkada earthquake event in Greece and the 2020 Mw 6.4 Puerto Rico earthquake event. Both events triggered hundreds of shallow landslides involving thin soil mantle and variably weathered and fractured regolith. For the Lefkada event, mapped landslide source areas and 3D volumes are used for the analysis. The shear strength estimates vary spatially across the study area, correlating with landslide frequency and field observation of material properties. The impact of DEM resolution on the model results is also investigated. The best-matching modeled landslides exhibit similar triggering depths and grid cells with comparable slope angles across varying resolutions of the DEM, with limited impact on back-calculated strength parameters for 5-m and 10-m resolutions. The Puerto Rico inventory consists of only 2D mapped polygons of the entire landslide area, which produces acceptable results albeit with greater uncertainties than the Lefkada example. In this analysis, the inversion of landslide data requires assumption of source area as a percentage of total landslide area, which along with the lack of volume data, results in greater uncertainties in the normal stress of back-calculated strength datapoints and shear strength.

Field observations and seismic surveys offered additional information from the Lefkada event for which the RESTAB inversion results for material strength can be assessed. Shear wave velocity collected from multi-channel analysis of surface wave (MASW) surveys and geological strength index (GSI)-based Hoek-Brown shear strength estimates from outcrop data describe the strength of rock masses that did not fail during the earthquake. In contrast, the back analysis from the earthquake-triggered landslides gives an estimate of strength at the landslide failure surface. A correlation is established between shear wave velocity and shear strength, as measured by MASW tests and back-calculated from the 2015 Lefkada earthquake-induced landslide inventory, respectively. A best-fit equation is derived to describe the correlation and can be used to estimate strength of landslides with only shear wave velocity profiles. The back-calculated landslide strength is compared to the GSI-based Hoek-Brown shear strength from outcrops, and it is found that the GSI-based Hoek-Brown shear strength is higher than the back-calculated strength due to the latter being derived from the failure surface of landslides and being lower than the strength of stable materials.

The developed pseudo-3D procedure in RESTAB is also applied to the forward prediction of earthquake-triggered landslides (i.e., RESTAB-prediction). Compared to commonly-used grid cell-based prediction methods, the pseudo-3D prediction methodology has the advantages of predicting individual landslide number, location, and geometry. Prediction results are evaluated by various metrics, including landslide number and area density, correctly predicted ratio, overlapped ratio, centroid distance, ground failure capture, and efficiency. RESTAB-prediction performs better than cell-based prediction methods in all applicable metrics except landslide area density. The effect of the selection of seismic displacement models on the prediction results for the 2015 Lefkada earthquake event is found to be comparable to strength parameter uncertainty.

A common shortcoming of regional forward prediction models is the overprediction of landslides, which may be improved by supplementing machine learning algorithms to model results. Overprediction occurs because seismic displacement models determine a critical slope angle above which all steeper slopes are predicted to fail, whereas in reality, only a portion of the slopes do fail do. Likely, lack of site-specific information hampers further improvement of purely mechanistic models that predict stability for some steep slopes compared to others, such as spatially-varying material properties or consideration of seismic site-effects on local topography. For example, RESTAB-prediction successfully predicts landslides that broadly overlap geospatially with most mapped landslides and also regions with high mapped landslide density. However, when considering landslide number or area density, landslides are overpredicted. The addition of machine learning classification models with RESTAB-prediction results is used to improve regional earthquake-induced landslide prediction. XGBoost is selected as the classification model and five parameters, including landslide volume, ratio of 3D factor of safety (FS) to 1D FS, aspect, triggering elevation, and topographic roughness, are used as training features. The integration of XGBoost with training data of individual landslides successfully excludes many small overpredicted landslides, improving landslide number by 38%. However, other metrics of success, such as the true positive rate and overlapped ratio, decrease. Using predicted landslide grid cells improves some results such as area density and false positive rate compared to using individual landslides as training data, because a grid cell model can exclude overpredicted landslide grid cells that are partial components of landslide areas. But a grid cell approach performs worse for landslide number, true positive rate, and correctly predicted ratio. Therefore, it is demonstrated that mechanistic prediction methodologies (i.e., RESTAB-prediction) integrated with machine learning can improve the prediction of regional earthquake-induced landslides, but the improvement depends on the type of metric used for success. This points to the need of having a clear objective and specific metrics for success prior to landslide prediction.