On two implicit issues in prediction modeling of landslide susceptibility

Abstract—this contribution argues the interpretation of predictionrate curves of landslide susceptibility and the corresponding prediction-pattern uncertainties. The focus is on two main issues in any kind of spatial prediction modeling: (1) isolating the meaningful parts of prediction-rate curves from a cost-benefit point of view and (2) comparing the qualities of prediction patterns obtained by different mathematical models and/or dissimilar spatial evidences. Mathematical models, methods, and databases generated prediction maps (we prefer the term prediction patterns) hard to evaluate due to inevitable relativity of measures, representations and confidence. This is a major problem, with modelling assumptions and justifications mostly ignored or poorly discussed. We consider prediction-rate curves, obtained by cross-validation, as standardized procedure. The curves are the result of cross-validating prediction patterns with the distribution of occurrences more recent than the ones used to generate the patterns. The two issues concern all types of modeling independently of algorithmic complexity or database formats. We examine an application example and its analytical strategy to point at resolving problems of pattern evaluation, comparison and uncertainty measure. The need becomes evident of collaborative efforts towards solutions analyzing a common multiformat database.