Role of hydrological model structure in the assimilation of soil moisture for streamflow prediction

Soil moisture data assimilation (SMDA) has found growing application in the hydrological research community because of readily available satellite-based soil moisture (SM) data. Several past studies have explored the capability to assimilate observed SM in hydrological model to enhance streamflow prediction. However, the impact of the conceptual hydrological model (CHM) structure on the SMDA in streamflow prediction has not been investigated yet. In this study, to understand the CHM structure's role, we used three different CHMs for the SMDA: dynamic Budyko (DB), Genie Rural a 4 parametres Journalier (GR4J), and Probability Distributed Model (PDM) model. The SM obtained from Global Land Data Assimilation System (GLDAS) was assimilated using Ensemble Kalman Filter (EnKF) for 43 Model Parameter Estimation Experiment (MOPEX) basins. The GR4J model was found to be performed best from calibration and validation results, and the DB model was found to be performed least. The results show that the performance of the DB model improved during assimilation compared to its open-loop version for all the 43 basins. However, deterioration in model performance was observed for the GR4J and PDM model during assimilation for all basins except a few. The assimilated model performance was evaluated in respect of Assimilation Efficiency (AE) and ranged from 2.56 to 81.24%, -70.9 to 17.65%, and -71.38 to 24.86%, respectively, for the DB, GR4J, and PDM model. Further, we hypothesized that if the GLDAS SM is better than the model-simulated SM, then improvement in the model performance is observed due to the SMDA. The coefficient of determination (r(2)) between the SM simulated by all the three models without assimilation and the GLDAS SM was found for all the basins. Results indicated that the GR4J and PDM model captured SM better than the DB model. Moreover, to strengthen this hypothesis, we run the GR4J and PDM model deterministically, considering the daily GLDAS SM as the initial condition for streamflow simulation instead of the SM simulated by these models for the basins where deterioration in model performance was observed. For all the basins, it is found that the GR4J and PDM model performance deteriorated. Even though the GLDAS SM is used as every day's initial condition to simulate the streamflow, no improvement of model performance is observed. In general, our study underlines the significance of the CHM structure in the SMDA exercise.