Item

Abstract

The aim of this work to illustrate a framework for flood early warning systems based on Bayesian forecast-decision theory. Bayesian decision approach provides a framework for issuing warnings under uncertainty and requires the definition of a predictive probability density of the future state and of a utility function that quantifies, often in economic terms, the consequences of the various actions to be taken. The rational choice is the one that minimize the expected value of losses. The approach is illustrated using one-day ahead warnings relying on severe precipitation forecasts issued by the Italian Civil Protection Department over a Weather Vigilance Zone of the national warning system in Italy. For this purpose, starting from Quantitative Precipitation Forecasts (QPFs), a post-processing technique to evaluate future rainfall conditional predictive distributions, namely the Univariate version of the Model conditional Processor (UMCP), was adopted. The obtained probabilistic predictive distribution is then used within a Bayesian decision scheme with different utility functions describing subjectively evaluated costs under the decision whether or not to issue a warning. Results show that post-processed forecasts provide better performance in terms of accuracy and reliability than ensemble QPFs, tend to correct bias and are generally less under-dispersive than raw forecasts for the investigated area. The system performances assuming a Bayesian utility function minimization criterion are then compared with the traditional approach for issuing warnings based on deterministic thresholds, demonstrating improvements in the identification of critical events and a sensible gain in terms of total cost from warnings.