Evaluating map quality and model performance through iterative statistics-based landslide susceptibility in eastern KY

Statistics-based landslide susceptibility modeling involves numerous and compounding choices of machine-learning techniques, variable inputs, and model performance metrics. Much of the success and practicality of these models hinges on commonly used geomorphic variables and the quality and type of a landslide inventory. We build on a two-step machine learning method, which includes a decision tree algorithm followed by multinomial logistic regression, by applying the approach to a new study area in eastern Kentucky, USA. Specific changes include testing geomorphic variables not previously considered in models for this region, how the landslide inventory is used to acquire geomorphic data, and compilation of binary data for both landslide and non-landslide areas. Tuning the balance of landslide and non-landslide data by adding more non-landslides accounts for uncommon landscape features and anthropogenic alterations. Our results show that adding geomorphic variables and changes in utilization of landslide inventories has a significant impact on both receiver operating curve-area under the curve (ROC-AUC) and map quality. Assessment of map quality is dependent on expert knowledge, map result distributions, and political boundary cohesion between previously modeled areas. A geographically distributed landslide inventory and an unbalanced binary data table of landslide and non-landslide variables generated the lowest ROC-AUC score (0.78) but the highest-quality map. Although model iterations for the goal of achieving higher model accuracy are worth time and effort, more variables do not necessarily improve model performance or map quality. Qualitative knowledge about the landscape is critical for reliable and desired map outcomes, hazard communication, and risk reduction.