EVALUATING REFINEMENTS IN STATISTICS-BASED LANDSLIDE SUSCEPTIBILITY MODELING THROUGH MODEL PERFORMANCE AND MAP QUALITY

Geomorphic and statistics-based landslide susceptibility modeling can involve endless choices of geomorphic variable inputs, machine-learning techniques, and model performance metrics. Much of the success and practicality of these models hinges on the quality of commonly used geomorphic variables such as elevation, slope angle, curvature, plan curvature, roughness, and aspect. In addition, the quality and type of a landslide inventory (points vs. polygons, for example) used to acquire geomorphic data is critical to the resulting model performance and applicability of susceptibility maps. We build on a two-step machine learning approach which includes a decision tree algorithm followed by multinomial logistic regression. We apply this approach to a new study area in eastern KY, using input variables previously not considered in models for this region. We change how the landslide inventory is used to acquire the variables, as well as refine the creation of a binary landslide data table of both landslide and non-landslide data.

We accomplish this through three model refinements: (1) testing geomorphic variables, including topographic wetness, flow accumulation, flow direction, proximity to streams, and proximity to roads, (2) testing the effectiveness of using geographically separate landslide inventories to acquire the geomorphic variables that are perhaps more representative of landslide occurrence in the entire study area, and (3) adding 120 more non-landslides to account for uncommon landscape features and anthropogenic alterations that could potentially skew model results.

We concluded that refinements of variable testing and changes in utilization of landslide inventories control a wide range of results in both ROC-AUC and map quality. A refinement iteration that used a geographically distributed landslide inventory and an unbalanced binary data table of landslide and non-landslide variables generated the lowest ROC-AUC score but the highest-quality map. Map quality is dependent on expert knowledge and the reliability of connecting with previously modeled areas. Although model refinements are worth time and effort, more data does not necessarily improve model performance or map quality. Qualitative knowledge about the landscape is critical for desired map outcomes, hazard communication, and risk reduction.