The use of global versus region-specific data for the prediction of co-seismic landslides

Abstract

Listen

Translate article

This study compares the impact of global and region-specific variables on the prediction of co-seismic landslides, investigating whether the incorporation of higher-resolution data into a global framework can improve the prediction outcome for regional landslide assessments. Based on a statistical landslide model, which was developed through logistic regression involving various global landslide inventories, the likelihood of a landslide occurring during an earthquake is estimated for a regional event. Using the recently updated landslide inventory of the 2016 MW7.8 Kaikōura earthquake (centroid of landslide source area), the prediction performance of the model is evaluated by individually replacing the global variables for slope, compound topographic index, lithology and land cover with New Zealand specific datasets that offer higher resolution and/or more accurate information. Statistical metrics are used to compare the changes in the prediction performance of each modification, while probability maps and binary prediction maps are evaluated to identify spatial differences. The results suggest that calculating the slope variable from a higher-resolution DEM (e.g., 25 m) improves the model's prediction performance without changing the model equation or replacing the other global variables. Other region-specific variables such as land cover show stronger correlation but require an adjustment of the model coefficients (e.g., regression analysis). As the region-specific datasets increase the landslide estimates, a sensible interpretation of the model results is important, including the identification of a probability threshold which allows for a binary prediction (landslide occurs versus landslide does not occur). Future research should investigate the role of global and region-specific datasets for other earthquake events and consider incorporating other variables (e.g., local slope relief). The findings demonstrate the limitations of global variables for regional landslide predictions and illustrate the impact as well as implications of region-specific datasets on the model outcome, providing a better understanding on how incorporating region-specific datasets into a global approach can improve the landslide prediction in areas where regional models are not available or where inventory data is not sufficient to develop a robust model.