Improving near‐real‐time coseismic landslide models: Lessons learned from the 2016 Kaikōura, New Zealand, earthquake

Abstract

The U.S. Geological Survey (USGS) is developing near‐real‐time global earthquake‐triggered‐landslide products to augment the USGS Prompt Assessment of Global Earthquakes for Response (PAGER) system. The 14 November 2016 M w "> M w Mw 7.8 Kaikōura, New Zealand, earthquake provided a test case for evaluating the performance and near‐real‐time response applicability of three published global seismically induced landslide models. All three models obtain shaking estimates from the USGS ShakeMap, which is updated and sometimes changes significantly in the hours to days after an earthquake. The Kaikōura earthquake is a particularly valuable event that helps us better understand how changes to the ShakeMap affect the landslide models because the ShakeMap evolved significantly over several weeks as multifault rupture and seismic data were incorporated. We used the detailed landslide inventory available for this event for qualitative landslide model evaluation. We found that once a point source was replaced with an approximate rupture extent in ShakeMap, the landslide models were all successful at roughly identifying the area of highest hazard. This is notable, given that the models are relatively simple, coarse in resolution, and are based solely on input proxies that are globally available. However, all of the models dramatically overpredicted the hazard level, which indicates that improvements can be made. Subsequent updates to the ShakeMap resulted in improvements to model performance by some metrics and declining performance by others. In all cases, details of the ShakeMap strongly controlled the spatial pattern, even when those details were erroneous, such as the inclusion of a fault segment that did not rupture. If maps of landslide hazard are to be used effectively for rapid response, then we need to understand and clearly communicate the control that ShakeMap has over the models and how that typically evolves with time and is (or is not) reflected in reported uncertainties.