Automated determination of landslide locations after large trigger events: advantages and disadvantages compared to manual mapping

Abstract

Earthquakes in mountainous areas can trigger thousands of co-seismic landslides, causing significant damage, hampering relief efforts, and rapidly redistributing sediment across the landscape. Efforts to understand the controls on these landslides rely heavily on manually mapped landslide inventories, but these are costly and time-consuming to collect, and their reproducibility is not typically well constrained. Here we develop a new automated landslide detection index (ALDI) algorithm based on pixel-wise normalised difference vegetation index (NDVI) differencing of Landsat time series within Google Earth Engine accounting for seasonality. We compare classified inventories to manually mapped inventories from five recent earthquakes: Kashmir in 2005, Aysén in 2007, Wenchuan in 2008, Haiti in 2010, and Gorkha in 2015. We test the ability of ALDI to recover landslide locations (using receiver operating characteristic – ROC – curves) and landslide sizes (in terms of landslide area–frequency statistics). We find that ALDI more skilfully identifies landslide locations than published inventories in 10 of 14 cases when ALDI is locally optimised and in 8 of 14 cases both when ALDI is globally optimised and in holdback testing. These results reflect not only good performance of the automated approach but also surprisingly poor performance of manual mapping, which has implications both for how future classifiers are tested and for the interpretations that are based on these inventories. We find that manual mapping, which typically uses finer-resolution imagery, more skilfully captures the landslide area–frequency statistics, likely due to reductions in both the censoring of individual small landslides and amalgamation of landslide clusters relative to ALDI. We conclude that ALDI is a viable alternative to manual mapping in terms of its ability to identify landslide-affected locations but is less suitable for detecting small isolated landslides or precise landslide geometry. Its fast run time, cost-free image requirements, and near-global coverage suggest the potential to significantly improve the coverage and quantity of landslide inventories. Furthermore, its simplicity (pixel-wise analysis only) and parsimony of inputs (optical imagery only) mean that considerable further improvement should be possible.