Improving Streamflow Estimates using an Efficient Time Variant Multivariate Assimilation of Soil Moisture and Streamflow Observations

During Data Assimilation (DA) in a hydrological model, observations of soil moisture (SM) and streamflow (Q) at interior locations are often assimilated together during the multivariate case to improve streamflow estimates at the catchment outlet. In addition to model states, model parameters need to be updated periodically to account for the variations caused by climatic and human factors during the assimilation period. Therefore, in this study, time-varying multivariate assimilation of ASCAT SM observations and streamflow gauge data from interior sites are ingested into a conceptual two-parameter model, which simulates streamflow using a water budget equation. The Bharathapuzha river basin, lying in the Western Ghats of Southern India is chosen as the study area. In this study, the Ensemble Kalman filter (EnKF), a sequential assimilation approach, is utilized to update the model’s states and parameters at a daily time step. Meanwhile, the computational burden of assimilating such a massive observation needs to be dealt with. A plausible solution is to perform assimilation only at those timesteps when the model is sensitive to the assimilating variable. Consequently, two assimilation scenarios were performed apart from the open-loop (OL) simulations. In the first scenario, all the available SM observations are assimilated irrespective of their sensitivity (DA1). Whereas, in the second scenario, only sensitive SM observations are assimilated into the model (DA2). Results revealed that during both the assimilation scenarios, the model showed improved performance as compared to the open-loop simulations. KGE value improved from 0.68 (during OL) to 0.85 (during DA1) and 0.81 (during DA2). An intriguing fact is that during the second scenario (DA2) when only a subset of sensitive observations was assimilated, the model still showed similar results as DA1. Results highlight that assimilating only spatiotemporally sensitive observations would not affect the model’s performance substantially. Instead, the assimilation efficiency can be enhanced by abbreviating the computational burden.