MAPPING LANDSLIDE SUSCEPTIBILITY OVER LARGE REGIONS WITH SPARSE DATA – A REALITY CHECK

Landslide susceptibility maps indicate the spatial distribution of future and existing landslide likelihood. Modeling landslide susceptibility over large and/or diverse terrains remains a challenge due to the sparsity of data and the variability in triggering conditions. Several approaches are used to mitigate these challenges including extending statistical models developed in data-rich regions to data-poor regions, training models on sparse but well-distributed landslide-inventory data, and restricting model training and application to regions with similar environmental attributes. However, to our knowledge, no study has attempted to systematically evaluate the effectiveness of these methods. Here we introduce a new statistical framework that facilitates comparisons between different statistical susceptibility models. We use this framework to evaluate a set of models trained on a variety of spatially isolated landslide inventories collected at sub-regional scales (~1000 km²) over diverse terrains across the United States. Results show significant variations in predictor effects (i.e., how a predictor value influences the susceptibility output) depending on the inventory used to train the models. The inconsistent predictor effects cause low accuracies when testing models on inventories outside the domain of the training data. Grouping test and training sets according to physiographic and ecological characteristics does not mitigate these challenges. We also show that using sparse training data distributed across the entire modeling domain can create susceptibility models that are relatively accurate when applied to the test data from the same domain. This study highlights limitations of present landslide inventories, and their spatial coverage.