Abstract
<strong class="journal-contentHeaderColor">Abstract.</strong> This study assesses global landslide susceptibility (LSS) at the coarse 36âkm spatial resolution of global satellite soil moisture observations to prepare for a subsequent combination of a global LSS map with dynamic satellite-based soil moisture estimates for landslide modeling. Global LSS estimation contains uncertainty, arising from errors in the underlying data, the spatial mismatch between landslide events and predictor information, and large-scale LSS model generalizations. For a reliable uncertainty assessment, this study combines methods from the landslide community with common practices in meteorological modeling to create an ensemble of global LSS maps. The predictive LSS models are obtained from a mixed effects logistic regression, associating hydrologically triggered landslide data from the Global Landslide Catalog (GLC) with predictor variables describing the landscape. The latter are taken from the Catchment land surface modeling system (including input parameters of soil (hydrological) properties and resulting climatological statistics of water budget estimates), as well as geomorphological and lithological data. Road network density is introduced as a random effect to mitigate potential landslide inventory bias. We use a blocked random cross validation to assess the <i>model uncertainty</i> that propagates into the LSS maps. To account for other uncertainty sources, such as <i>input uncertainty</i>, we also perturb the predictor variables and obtain an ensemble of LSS maps. The perturbations are optimized so that the <i>total predicted uncertainty</i> fits the observed discrepancy between the ensemble average LSS and the landslide presence or absence from the GLC. We find that the most reliable <i>total uncertainty</i> estimates are obtained through the inclusion of a topography-dependent perturbation between 15â% and 20â% to the predictor variables. The areas with the largest LSS uncertainty coincide with moderate ensemble average LSS, because of the asymptotic nature of the LSS model. The spatial patterns of the average LSS agree well with previous global studies and yield areas under the receiver operating characteristic between 0.84 and 0.92 for independent regional to continental landslide inventories.
Highlights
Mitigating landslide impacts requires a good understanding of the spatial and temporal patterns of landslide occurrence
We summarize its information by the area under the Receiver Operation Characteristic (ROC) curve (AUC)
This study presents the first global landslide susceptibility (LSS) map directly developed to be compatible with satellite soil moisture products retrieved from passive microwave sensors, i.e. at a spatial resolution of 36 km
Summary
Mitigating landslide impacts requires a good understanding of the spatial and temporal patterns of landslide occurrence. The total ensemble uncertainty, resulting from the combination of these methods that account for model and input uncer tainty respectively, is assumed to be reliable if it matches the observed ‘actual’ total uncertainty The latter is estimated by comparing the predicted average LSS against the observed presence and absence of landslides. To limit biases from unreliable and confounding definitions of landslide absence grid cells for the model creation, we introduce a novel approach based on a ‘characteristic distance’ between landslides After having taken these steps to limit the introduced uncertainty, the B-CV is used to introduce model uncertainty and we further add (and tune) ensemble perturbations to the selected predictor variables to obtain a reliable total ensemble uncertainty.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
