Identifying time-varying hydrological model parameters to improve simulation efficiency by the ensemble Kalman filter: A joint assimilation of streamflow and actual evapotranspiration

Abstract

Hydrological model parameters are essential for model simulation, which may vary with time owing to climatic variations and human activities. As a result, the implementation of stationary parameters may lead to inaccurate streamflow simulation. Actual evapotranspiration (ET) is a crucial component of hydrological cycles with an important influence on regional climate characteristics. Our aim was to investigate an appropriate way to estimate time-varying parameters using streamflow and ET observations, and to explore the how the simulation efficiency would change when ET data were added into assimilation. Data assimilation techniques can automatically adjust hydrological model parameters according to changing conditions. The ensemble Kalman filter (EnKF) technique was used to assimilate observations into a two-parameter monthly water balance model. Four data assimilation schemes were designed and applied in a synthetic experiment and 173 catchments in USA, including: (1) sole assimilation of streamflow to update both parameters, (2) joint assimilation of streamflow and ET to simultaneously update parameters, (3) sole assimilation of ET first to update parameter C (measuring evapotranspiration), and then subsequently joint assimilation of streamflow and ET to update parameter SC (measuring water storage capacity), and (4) sole assimilation of ET to update parameter C, and then sole assimilation of streamflow to update parameter SC. The four schemes were compared in terms of deterministic and probabilistic model performances, and model parameter correlations. The results indicated that: (1) the estimation of parameter C can be improved by assimilating ET observations into the model, and the time-varying model parameters obtained by adding ET data into assimilation led to a significant enhancement in deterministic streamflow and ET prediction, (2) among the three joint assimilation schemes, the deterministic model simulation performances were similar, while the ensemble predictions were more reliable in scheme 2. The joint assimilation of streamflow and ET to update both parameters in one step, i.e., scheme 2, outperformed the other two schemes in parameter SC estimation and model performance, (3) the ability to enhance simulation accuracy and to weaken correlations between parameters was more significant in humid catchments. The joint assimilation was helpful for the application of multiple sources of information into a hydrological simulation.