National-scale shallow landslide precipitation thresholds in Austria for early warning: A comparison of two modelling approaches

To improve and automate the shallow landslide component of the already operating Austrian early warning system AMAS (Austrian Multi-Hazard Impact-based Advice Service), regional precipitation thresholds are needed.  Both the existing warning system and the precipitation thresholds developed in this study do not target individual landslides, but focus on severe regional events involving multiple landslides. Here, we present preliminary results of precipitation threshold modelling at national scale. 

Historic regional events were extracted from Austria-wide landslide inventories, including GEORIOS (GeoSphere Austria), the WLK database of the Austrian Torrent and Avalanche Control, as well as landslide inventories from different Austrian federal states. Landslide absence observations were identified by selecting landslide-free precipitation events with more than 20 mm in 24 h for which no indications of landslides were found after screening additional sources such as fire brigade reports, police records, local authorities, and VIOLA – the severe weather database of GeoSphere Austria.

Taking into account the diverse environmental conditions under which landslides occur, Austria was divided into 21 geo-climatic regions using hierarchical cluster analysis, which considered geological, geomorphological, pedological and climatic factors, complemented by expert knowledge. While our aim is to model the precipitation thresholds for each of the 21 geo-climatic regions in Austria, we present preliminary results for two study areas of the Fischbacher Alps and the Vorarlberger Molasse. Lessons learned in these areas will be applied to nationwide modeling.

Precipitation threshold modeling was performed using two different techniques: (i) a data-driven approach based on generalized additive models (GAMs), which combines triggering and antecedent precipitation, and (ii) a quantile regression approach, which defines the onset of relevant precipitation following a dry period. For both approaches, precipitation data from INCA (Integrated Nowcasting through Comprehensive Analysis, combined radar and station data) were used with hourly resolution.  

To optimize the results, the durations of triggering and antecedent precipitation in the GAM model, as well as the dry-period duration and the maximum precipitation threshold during the dry period in the quantile regression model, are systematically varied. Additional model variants consider the inclusion of the antecedent precipitation index (API) and the use of different landslide samples (e.g., representatively sampled points across different rainfall events) for both models. The best modeling results are selected via ROC-based cross-validation complemented with expert plausibility checks (e.g., longer antecedent precipitation for fine-grained soils). 

First GAM results showed very high predictive performance, with mean cross-validation AUROCs exceeding 0.9. Including a third variable in the model, namely peak 1-hour rainfall within the triggering window, alongside cumulative triggering and antecedent precipitation further improved the model, and the modeled relationships appeared plausible. Early quantile-regression estimates of intensity-duration (ID) thresholds are consistent with prior work (e.g., Guzzetti et al., 2008; Marra et al., 2014) but exhibit a steeper power-law decay. These results are sensitive to event-sample representativeness as well as the delineation of triggering rainfall, and they reveal spatial heterogeneity consistent with differing geological and meteorological predisposition.