The necessity to consider the landslide data origin in statistically-based spatial predictive modelling – A landslide intervention index for South Tyrol (Italy)

Most statistically-based landslide susceptibility maps are supposed to portray the relative likelihood of an area to be affected by future landslides. Literature indicates that vital modelling decisions, such as the selection of explanatory variables, are frequently based on quantitative criteria (e.g. predictive performance). The results obtained by apparently well-performing statistical models are also used to infer the causes of slope instability and to identify landslide “safe” terrain. It seems that comparably few studies pay particular attention to background information associated with the available landslide data. This research hypothesizes that inappropriate modelling decisions and wrong conclusions are likely to follow whenever the origin of the underlying landslide data is ignored. The aims were to (i) analyze the South Tyrolean landslide inventory in the context of its origin in order to (ii) highlight potential pitfalls of performance driven procedures and to (iii) develop a predictive model that takes landslide background information into account. The available landslide data (1928 slide-type movements) of the province of South Tyrol (~7400 km²) consists of positionally accurate points that depict the scarp location of events that induced interventions by e.g. the road service or the geological office. An initial exploratory statistical analysis revealed general relationships between landslide presence/absence data and frequently used explanatory variables. Subsequent modelling was based on a Generalized Additive Mixed Effects Model that allowed accounting for (non-linear) fixed effects and additional “nuisance” variables (random intercepts). The evaluation of the models (diverse variable combinations) focused on modelled relationships, variable importance, spatial and non-spatial predictive performance and the final prediction surfaces. The results highlighted that the best performing models did not reflect the “actual” landslide susceptibility situation. A critical interpretation led to the conclusion that the models simultaneously reflected both, effects likely related to slope instability (e.g. low likelihood of flat and very steep terrain) and effects rather associated with the provincial landslide intervention strategy (e.g. few interventions at high altitudes, increasing number of interventions with decreasing distance to infrastructure). Attempts to separate the nuisance related to “intervention effects” from the actual landslide effects using mixed effects modelling proved to be challenging, also due to omnipresent spatial interrelations among the explanatory variables and the fact that some variables concurrently represent effects related to landslide predisposition and effects associated with the intervention strategy (e.g. altitude). We developed a well-performing predictive landslide intervention index that is in line with the actual data origin and allows identifying areas where future interventions are more or less likely to take place. The efficiency of past interventions (e.g. stabilization of slopes) was demonstrated during recent storm events, because previously stabilized slopes were not affected by new landslides. This also showed that the correct interpretation of the final map requires a simultaneous visualization of both, the spatially predicted index (from low to high) and the available landslide inventory (low likelihood due to past interventions). The results confirm that wrong conclusions can be drawn from excellently performing statistical models whenever qualitative background information is disregarded.