Spatio‐Temporal Model Variance‐Covariance Approach to Assimilating Streamflow Observations Into a Distributed Landscape Water Balance Model

Abstract

AbstractAccurate initial conditions play a critical role in improving predictive accuracy of hydrological models for quantities such as streamflow generation. Streamflow observations from in situ gauging sites have been assimilated in a wide range of past research to improve lumped catchment streamflow. However, spatio‐temporal state updating in distributed hydrological models through streamflow data assimilation (DA) remains a challenge due to the large dimensional disparity between model state space and observation space. This study explores the use of model spatio‐temporal variance‐covariance from simulated climatology as a surrogate for model error, as a way of distributing constraint from gauge measurements spatially. In the approach developed here, grid cells within and surrounding gauged catchments were updated simultaneously. The proposed DA method leads to a substantial improvement in performance and reliability compared to open‐loop (OL) estimates of streamflow. In particular, for “ungauged” catchments updated using streamflow observations from neighboring catchments, the mean absolute error (MAE) was reduced by 20% on average after the assimilation. In addition, improvements from using the analyzed states as initial conditions were found to persist for more than 1 week, and even longer for catchments with high annual runoff. The improvement in streamflow forecasts using initial conditions from DA can be more than 40% in terms of MAE compared to OL results for 1 day lead time, and more than 10% for 5 days lead time.