Machine learning per l'idrologia: applicazione del metodo random forests per la previsione degli eventi di piena fluviale con un approccio probabilistico

Machine Learning techniques allow to maximize the use of information in real time. The Random Forests method can be counted among the most recent and performing Machine Learning techniques. Taking advantage of the potential of this method, this PhD thesis faces two different case studies and two different forecast models have been elaborated. The first case study focused on the main rivers of the Emilia-Romagna Region, characterized by very short response times. The choice of these rivers arise from the several flood events in these basins, happened in recent years, mostly of the "flash flood" type. The second case study concerns the main sections of the Po River, where the propagation time of the flood is greater than the water courses, if compared to the first case study. Starting from a large amount of data and according to the objectives to achieve, for both case studies the first step was to select and define the input data. For the elaboration of the model to the rivers of Emilia-Romagna, the observed data have been considered. Instead, for the Po River basin, the observed data and the forecast data from the Mike11 NAM / HD model chain were exploited. Taking advantage of one of the main characteristics of the Random Forests method, a probability of occurrence has been estimated: this information is suitable both in the technical phase and in the decision-making phase for any civil protection and intervention activities. All the processing, the used data and the developed models were performed in the R environment. At the end of the validation phase, the encouraging results allowed to insert the developed model in the first case study in the operational architecture of FEWS (Flood Early Warning System).