Assimilating in situ and radar altimetry data into a large-scale hydrologic-hydrodynamic model for streamflow forecast in the Amazon River basin

Large-scale hydrological models and forecast systems may be important tools to reduce the vulnerability of local population in places such as the Amazon River basin, where extreme hydrological events have occurred in the past few years. Due to the size of the basin and the slow speed of movement of its floods, uncertainty on model initial conditions (ICs) may play an important role for discharge forecasts using large scale hydrological models, even for relatively large lead times (~ 1 to 3 months). Data assimilation (DA) methods may provide an interesting way of merging both in situ and newly remotely sensed observations with models to estimate optimal ICs. We present the development and evaluation of a data assimilation framework for both gauged and radar altimetry based discharge and water levels into a large scale hydrological-hydrodynamic model of the Amazon River basin. We also explore the usefulness of such system to provide streamflow forecasts when forced by past climate and based mostly on model initial conditions. This work is in the context of recent developments of techniques for integrating information from models and remotely sensed data, and also of regional/global hydrological forecast systems including poorly gauged basins. We use the conceptual and physically based MGB-IPH model. The model uses the Penman Monteith for evapotranspiration and the Moore and Clarke model for soil water storage. River dynamics is simulated using full Saint-Venant equations and a simple floodplain storage model. The model was forced using satellite-derived daily rainfall (TRMM 3B42). We implemented a DA scheme based on the Ensemble Kalman Filter (EnKF) capable of assimilating three types of data: (1) discharge observations; (2) water levels provided by the ENVISAT radar altimeter; and (3) discharge estimated from radar altimetry. All state variables of the hydrological model were updated at each analyses time step. Model state variables errors were generated by perturbing the precipitation forcing field from the original value, using log-normally distributed, time and spatially correlated error fields. Streamflow forecasts were generated using an ensemble streamflow prediction approach, where the model starts with initial conditions obtained from the data assimilation step and where the model is forced by an ensemble of observed precipitation time series resampled from past years. First experiments show that an EnKF based assimilation of discharge observations and remotely sensed water levels improve estimates of the large scale hydrologic-hydrodynamic model in the Amazon River basin. Results improve mostly in large tributaries and in some cases errors increase at smaller rivers. Assimilation of water levels can also improve discharge estimates at gauges located near ENVISAT altimetry stations. We show the evaluation of the DA scheme using different configurations and kinds of observations. Finally, we evaluate the accuracy of streamflow forecasts at different spatial scales and lead times.