Comparing methods for determining landslide early warning thresholds: potential use of non-triggering rainfall for locations with scarce landslide data availability

Rainfall intensity-duration landslide-triggering thresholds have become widespread for the development of landslide early warning systems. Thresholds can be in principle determined using rainfall event datasets of three types: (a) rainfall events associated with landslides (triggering rainfall) only, (b) rainfall events not associated with landslides (non-triggering rainfall) only, (c) both triggering and non-triggering rainfall. In this paper, through Monte Carlo simulation, we compare these three possible approaches based on the following statistical properties: robustness, sampling variation, and performance. It is found that methods based only on triggering rainfall can be the worst with respect to those three investigated properties. Methods based on both triggering and non-triggering rainfall perform the best, as they could be built to provide the best trade-off between correct and wrong predictions; they are also robust, but still require a quite large sample to sufficiently limit the sampling variation of the threshold parameters. On the other side, methods based on non-triggering rainfall only, which are mostly overlooked in the literature, imply good robustness and low sampling variation, and performances that can often be acceptable and better than thresholds derived from only triggering events. To use solely triggering rainfall-which is the most common practice in the literature-yields to thresholds with the worse statistical properties, except when there is a clear separation between triggering and non-triggering events. Based on these results, it can be stated that methods based only on non-triggering rainfall deserve wider attention. Methods for threshold identification based on only non-triggering rainfall may have the practical advantage that can be in principle used where limited information on landslide occurrence is available (newly instrumented areas). The fact that relatively large samples (about 200 landslides events) are needed for a sufficiently precise estimation of threshold parameters when using triggering rainfall suggests that threshold determination in future applications may start from identifying thresholds from non-triggering events only, and then move to methods considering also the triggering events as landslide information starts to become more available.