A new observational-modeling framework for flash-flood forecasting in complex-terrain watersheds.

The watershed determined by Aburrá Valley system, located in northwestern Colombia, has significant urban development and steep hills. These features, together with the typical intense storms of the region, make the watershed prone to the occurrence of flash floods during the rainy seasons, affecting vulnerable communities. We propose a hybrid observational-modeling strategy to generate 30-minute discharge forecasts in different locations of the watershed, using an operational distributed hydrological model, information from stream gauges, and weather radar-derived precipitation using a quantitative precipitation estimation (QPE) technique. The forecast methodology is triggered when any stream gauge of interest reports levels over a predefined threshold. As a first step, the model uses different rainfall scenarios for the following 30 minutes. Every 5 minutes, the model forecast is executed after updating the observed rainfall and the rainfall scenarios. The scenarios correspond to (i) a lagrangian extrapolation of the precipitation fields, (ii) to a cellular automata-based extrapolation and to (iii) the last observed rain field multiplied by a time-varying ad-hoc factor based on historical event analysis. To parametrize the hydrological model and to validate the prediction methodology, we use 173 storm events from 2013 to 2018. The methodology is evaluated using the Nash coefficient, the Klin-Gupta index, differences in time-to-peak discharge, peak-discharge differences, and total storm-event volume differences. Operationally, the forecasted streamflow corresponds to the scenario with the best historical performance, given the total amount of observed rainfall. The overall results suggest that the described approach is promising. However, there are still some cases in which the method leads to discharge underestimation. Considering the forecast uncertainty, the results show that it is possible to design flash floods alerts using this simple but robust methodology.