Nele Kristin Meyer, Wolfgang Schwanghart, Oliver Korup, Bard Romstad, Bernd Etzelmuller

• The Norwegian traffic network is impacted by about 2000 landslides, avalanches, and debris flows each year that incur high economic losses. Despite the urgent need to mitigate future losses, efforts to locate potential debris flow source areas have been rare at the regional scale. We tackle this research gap by exploring a minimal set of possible topographic predictors of debris flow initiation that we input to a Weights-of-Evidence (WofE) model for mapping the regional susceptibility to debris flows in western Norway. We use an inventory of 429 debris flows that were recorded between 1979 and 2008, and use the terrain variables of slope, total curvature, and contributing area (flow accumulation) to compute the posterior probabilities of local debris flow occurrence. The novelty of our approach is that we quantify the uncertainties in the WofE approach arising from different predictor classification schemes and data input, while estimating model accuracy and predictive performance from independent test data. Our results show that aThe Norwegian traffic network is impacted by about 2000 landslides, avalanches, and debris flows each year that incur high economic losses. Despite the urgent need to mitigate future losses, efforts to locate potential debris flow source areas have been rare at the regional scale. We tackle this research gap by exploring a minimal set of possible topographic predictors of debris flow initiation that we input to a Weights-of-Evidence (WofE) model for mapping the regional susceptibility to debris flows in western Norway. We use an inventory of 429 debris flows that were recorded between 1979 and 2008, and use the terrain variables of slope, total curvature, and contributing area (flow accumulation) to compute the posterior probabilities of local debris flow occurrence. The novelty of our approach is that we quantify the uncertainties in the WofE approach arising from different predictor classification schemes and data input, while estimating model accuracy and predictive performance from independent test data. Our results show that a percentile-based classification scheme excels over a manual classification of the predictor variables because differing abundances in manually defined bins reduce the reliability of the conditional independence tests, a key, and often neglected, prerequisite for the WofE method. The conditional dependence between total curvature and flow accumulation precludes their joint use in the model. Slope gradient has the highest true positive rate (88%), although the fraction of area classified as susceptible is very large (37%). The predictive performance, i.e. the reduction of false positives, is improved when combined with either total curvature or flow accumulation. Bootstrapping shows that the combination of slope and flow accumulation provides more reliable predictions than the combination of slope and total curvature, and helps refining the use of slope-area plots for identifying morphometric fingerprints of debris flow source areas, an approach used outside the field of landslide susceptibility assessments.…