Variational assimilation of streamflow data in distributed flood forecasting

Data assimilation has the potential to improve flood forecasting. However, research efforts are still needed for an effective development of assimilation schemes suitable for operational usage, especially in case of distributed hydrologic models. This work presents a new assimilation system of streamflow data from multiple locations in a distributed hydrologic model. The system adopts a mixed variational-Monte Carlo approach, and is here tested with the hydrologic model MOBIDIC, that is part of the operational flood forecasting chain for Arno river in central Italy. The main objective of the work is to evaluate the actual gain that the system can lead to flood predictions in a real-time operational usage. Accordingly, a specifically designed assessment strategy is employed. It is based on several hindcast experiments that include both high flow and false alarm events in the period 2009–2014 in Arno river basin. Results show that the assimilation system can significantly increase the accuracy of flow predictions in respect to open loop simulations in both cases. Specific performances depend on location and event, but in the majority of cases the error on predicted peak flow is reduced of more than 50% with a lead time of around 10 h. The analysis reveals also that the structure of the hydrologic model, the coherence between observations at various sites, and the initial watershed saturation level, considerably affect the obtainable performances. Conditions that may lead to a worsening of open loop predictions are identified and discussed.