1. Data-driven regional frequency analysis of sub-daily rainfall extremes across large and morpho-climatically heterogeneous regions
- Author
-
Magnini, A., Lombardi, M., Ouarda, T., and Castellarin, A.
- Abstract
The scientific literature reports on various approaches to regional frequency analysis (RFA) of rainfall extremes. Traditional regional models usually refer to a single time-aggregation interval. They are developed for medium-to-small and climatically and morphologically homogeneous regions, which leads to higher accuracy, yet limits the models' exportability. We propose a different and innovative approach to RFA of rainfall extremes; it consists of ensembles of unsupervised artificial neural networks (ANNs), predicting the parameters of a Gumbel distribution for dimensionless (i.e. standardized by the local value of the mean annual maximum rainfall depth) annual maximum rainfall depths at any location in the study area and for any time-aggregation interval in the 1-24 hours range. Predictions are based on several morpho-climatic descriptors, used as covariates of precipitation extremes, and the applicability of the models extends to a large and morphologically as well as climatically heterogeneous region (Western and Central Northern Italy). ANNs are trained on over 2300 Annual Maximum Series of rainfall depths collected between 1928 and 2011 and associated with five hourly time-aggregation intervals. Based on a comprehensive analysis of predictions obtained at 100 randomly selected validation points, we discuss the potential and drawbacks of ensembles of unsupervised ANNs for integrated (i.e., across various time-aggregation intervals) and multivariate (i.e., considering numerous morphoclimatic covariates) RFA of sub-daily rainfall extremes., The 28th IUGG General Assembly (IUGG2023) (Berlin 2023)
- Published
- 2023
- Full Text
- View/download PDF