LANDSLIDE SUSCEPTIBILITY MODELLING OF ALBERTA, CANADA: COMPARATIVE RESULTS FROM MULTIPLE STATISTICAL AND MACHINE-LEARNING PREDICTION METHOD

Landslides represent one of the most widespread geological hazards and globally account for significant damages to property and infrastructure. This can render land unsuitable for development without mitigative measures. Spatial analysis techniques that predict the degree to which terrain is predisposed to developing landslides, known as landslide susceptibility models (LSMs), represent a widely-applied method for delineating potentially landslide prone terrain at a regional-scale. This study focuses on the development of a data-driven, statistical-type application for regional-scale landslide susceptibility modelling of dominantly slow- to extremely-slow moving landslides across the plains and shield regions of Alberta, Canada. A significant difference between this study and previous studies is the size of the model region, which covers an area of 606,443 km² and includes a heterogeneous range of geological, topographic and climatic settings. A rigorous evaluation of multiple machine learning techniques has not been performed before at this scale. We have assessed the predictive performances of the widely applied logistic regression against some of the most commonly used and/or novel machine learning classifiers, and examined the impact of hyperparameter tuning in the modelling process, a step which has been ignored in most previous studies. Further, we investigated the sensitivity of each classifier to additional factors, including the size of the available training dataset, and the relative dependency on a wide range of predictors in a region that is typical of a continental Interior Plains setting. Strong quantitative and visible geomorphic differences were observed between the resulting susceptibility predictions, demonstrating the advantages of assessing of the results obtained using both statistical and machine-learning techniques.