A Bayesian framework for the multiscale assessment of storm severity and related uncertainties

Abstract

This article proposes a statistical framework for assessing the multiscale severity of a given storm at a given location. By severity we refer to the rareness of the storm event, as measured by the return period. Rather than focusing on predetermined spatiotemporal scales, we consider a model assessing the return period of a storm event observed across the continuum of durations and areas around a focus location. We develop a Bayesian intensity–duration–area–frequency model based on extreme value distribution and space–time scale invariance hypotheses. The model allows us to derive an analytical expression of the return period for any duration and area, while the Bayesian framework allows us by construction to assess the related uncertainties. We apply this framework to high-resolution radar–rain gauge reanalysis data covering a mountainous region of southern France during the autumns 2008–15 and comprising 50 rain events. We estimate the model at two grid points located a few kilometers apart on either side of the mountain crest, considering spatiotemporal scales ranging over 3–48 h and 1–2025 km2. We show that at all scales and for all significant events, the return period uncertainties are skewed to the right, evidencing the need of considering uncertainty to avoid systematic risk underestimation. We also reveal the large variability of the storm severity both at short distance and across scales, due to both the natural variability of rainfall and the mask effect induced by the mountain crest.