Flood Alert System For Early Warning In Mountainous Coastal Watersheds: Coupling Data-Driven And Physically Based Hydrological Models

Early warning systems are an essential tool for preventing flood risks and managing alert situations from real time data-driven outputs. In Mediterranean coastal regions, the response time is usually a constraint for decision-making since many watersheds have travel times of just a few hours from head sources to the floodplain in the valleys. In such situations, the anticipation of the alert situation in terms of probability occurrence is a key fact for managing alert conditions and mitigate flood consequences. Coupling data-driven forecasting with hydrological models may constitute a robust tool for alert assessment, provided that sound calibrations are available. This work shows the implementation of a graphic user interface (GUI) developed in MATLAB intended to serve for flooding forecast managers. The software is formed by a very straightforward group of windows that present schematic information related to a very specific and intuitive value of warning risk. This simple front-end encloses two complex modeling techniques for forecasting somehow hidden for the final user. On one hand, there is a data driven set of pre-trained models based on machine learning techniques (Artificial Neural Networks, Radial Basis Functions, Bayesian Neural Networks, and Gaussian Processes), which are capable of making a quick forecast of key variables at certain control points, with an associated uncertainty level. On the other hand, when the warning risk is above a given threshold, a cascade of physically-based models can be launched to perform a more detailed forecast: a sequence of weather forecasting from mesoscale models, which feed the physically based and fully distributed hydrological model WiMMed, and a hydraulic 1D model for flood forecast (WinForCe), both programmed in C++ and run in command line mode. Results are presented in terms of flooded area, probability, damage, and cost to quantify a final risk value already defined for statistical modelling