Reducing subjectivity in hydrometeorological thresholds for landslide early warning systems

Abstract:

Landslide warning systems are a common and effective tool for risk reduction because they can provide advanced notice of the increased potential for shallow landsliding. A primary challenge for informing these systems is defining thresholds that minimize both the number of failed and false alarms. Failed alarms do not allow effective actions to reduce the impact of unexpected landsliding, and excessive false alarms can erode public trust in landslide forecasts. Such thresholds have traditionally relied solely on rainfall intensity and duration, but a growing body of literature suggests that the number of errant warnings can be reduced through the simultaneous use of hydrological and meteorological conditions to forecast landslide initiation events. Hydrometeorological thresholds can employ soil moisture to account for antecedent conditions and recent or imminent rainfall amounts to account for the triggering event. Several methods have been developed to objectively optimize performance of both types of thresholds. However, several subjective decisions are involved in selecting the timescales for hydrometeorological threshold variables, the formulations for threshold equations, and the skill statistics for their optimization; the potential impacts related to such choices have not been fully explored. We use a new code for automated, objective optimization of hydrometeorological thresholds, HydroMet, to quantitatively evaluate these subjective factors for several monitoring sites across the U.S. We also explore whether using measured occurrences of positive pore-water pressures could serve as a conservative proxy for landslide occurrence to reduce failed alarms in cases where landslide timing is poorly constrained or unknown. Finally, we examine ways to account for uncertainty in rainfall forecasts when optimizing threshold performance. Our results show that the optimal skill statistic for balancing failed and false alarms can vary depending on the timescale used to define antecedent wetness and rainfall totals. Limiting the number of subjective decisions and quantifying their impact can provide insights into underlying triggering processes, and also improve warning accuracy, repeatability, and consistency to aid decision-making criteria for reducing landslide related losses.