DEVELOPMENT OF REGIONAL LANDSLIDE SUSCEPTIBILITY MODELS: A FIRST STEP TOWARDS MODEL TRANSFERABILITY

Landslides are a globally pervasive problem with the potential to cause significant fatalities and economic losses. Although landslides are widespread, many at-risk regions may lack the high-quality data and/or resources used in most landslide susceptibility analyses. This study aims to develop versatile regional susceptibility relationships using publicly available data and open-source software. We developed logistic regression and frequency ratio susceptibility relationships in 9 areas using different subsets of landslide data, resulting in 23 unique scenarios within the United States (where a scenario refers to a unique area and data combination). Areas were diverse in their geology, morphology, climate, and nature and quality of available landslide data. The transferability of select models to areas uninvolved in model development was also tested. The transferred models were trained using data from a single scenario (single-scenario cross-validation) or a combination of scenarios (multi-scenario cross-validation). We derived potential landslide contributing factors from a globally available digital surface model and used publicly available landslide inventories from state geological surveys. The contributing factors considered were elevation, slope, aspect, planform curvature, profile curvature, and topographic position index. We assessed model performance using the receiver operating characteristics area under the curve (AUC), where this curve represents the true positive proportion versus the false positive proportion over all possible diagnostic cutoff values. Models developed using high-quality landslide data delineating scarps, flanks, and individual slope movements performed very well (AUC 0.764 - 0.895). Models developed using landslide data dominated by deposits performed less well, but at or near an acceptable level (AUC 0.67 – 0.81). Models developed using older, lower quality landslide data did not perform at an acceptable level (AUC 0.63 – 0.64). The results of testing model transferability had acceptable results for some but not all regions (AUC 0.563 - 0.844). This study is a promising first step in developing generalized landslide susceptibility relationships that can be used in areas that share similar regional scale attributes.