Historically, hydrological models have been developed to represent land-atmosphere interactions by simulating water storage and water fluxes. These models, however, have their own unique characteristics (strength and weakness) in capturing different aspects of the water cycle, and their results are typically compared to or calibrated against in-situ observations such as river runoff measurements. As a result, there may be gross inaccuracies in the estimation of water storage states produced by these models. In this study, we present the novel approach of Dynamic Model Data Averaging (DMDA), which can be used to compare and merge multi-model water storage simulations with monthly Terrestrial Water Storage (TWS, a vertical summation of surface and sub-surface water storage) estimates from the Gravity Recovery And Climate Experiment (GRACE) satellite mission. Here, the main hypothesis is that merging GRACE data with multi-model outputs likely provides more skillful hydrological estimations compared to a single model or data set. Theoretically, the proposed DMDA combines the benefits of the Kalman Filter (KF) and Bayesian Model Averaging (BMA) techniques and has the capability to deal with various observations and models with different error structures. Based on the Bayes theory, DMDA provides time-variable weights for hydrological models to compute an average of their outputs that are best fited to GRACE TWS estimates. Numerically, the DMDA method is implemented by integrating the output of six hydrological and land surface models (PCR-GLOBWB, SURFEX-TRIP, LISFLOOD, HBV-SIMREG, W3RA, and ORCHIDEE) and monthly GRACE TWS estimates (2002–2012) within the world’s 33 largest river basins, while considering the inherent uncertainties of all inputs. Our results indicate that DMDA correctly separates GRACE TWS estimates into surface water, soil moisture and groundwater compartments. Linear trends fitted to the DMDA-derived groundwater compartment are found to be different from those of original models. This means that anthropogenic influences within the GRACE data, which are not well reflected by models, are introduced by DMDA. We also find that temporal correlation coefficients between the DMDA-derived individual water storage estimations (surface water, soil moisture, and groundwater) and the El Niño Southern Oscillation (ENSO) index are considerably increased compared to those derived between individual model simulations and ENSO (e.g., an increase from −0.2 to 0.6 in the Murray River Basin). For the Nile River Basin, they changed from 0.1 to 0.4 for the soil moisture, and from 0.3 to 0.7 for the surface water compartment. Comparisons between the DMDA-derived surface water and those from independent satellite altimetry observations indicate that after implementing DMDA, temporal correlation coefficients within major lakes are increased. Based on these results, we have gained confidence in the DMDA water storage estimates to be used for improving the characterization of water storage over broad regions of the globe.
- terrestrial water storage (TWS)
- dynamic model data averaging (DMDA)
- Kalman filter (KF)
- Bayesian model averaging (BMA)
- multi-hydrological models
- satellite altimetry
Mehrnegar, N., Jones, O., Bliss Singer, M., Schumacher, M., Bates, P., & Forootan, E. (2020). Comparing Global Hydrological Models and Combining them with GRACE data by Dynamic Model Data Averaging (DMDA). Advances in Water Resources, 138, . https://doi.org/10.1016/j.advwatres.2020.103528