Scrutiny of Appropriate Model Error Specification in Multivariate Assimilation Framework using mHM

Abstract:

Reliable and accurate predictions of regional scale water fluxes and states is of great challenge to the scientific community.
Several sectors of society (municipalities, agriculture, energy, etc.) may benefit from successful solutions to appropriately
quantify uncertainties in hydro-meteorological prediction systems, with particular attention to extreme weather conditions.
Increased availability and quality of near real-time data enables better understanding of predictive skill of forecasting
frameworks. To address this issue, automatic model-observation integrations are required for appropriate model
initializations. In this study, the effects of noise specification on the quality of hydrological forecasts is scrutinized via a
data assimilation system. This framework has been developed by incorporating the mesoscale hydrologic model (mHM,
{http://www.ufz.de/mhm) with particle filtering (PF) approach used for model state updating. In comparison with previous
works, lag PF is considered to better account for the response times of internal hydrologic processes.
The objective of this study is to assess the benefits of model state updating for prediction of water fluxes and states up to
3-month ahead forecast using particle filtering. The efficiency of this system is demonstrated in 10~large European basins.
We evaluate the model skill for five assimilation scenarios using observed
(1) discharge (Q); (2) MODIS evapotranspiration (ET); (3) GRACE terrestrial total water storage (TWS) anomaly; (4) ESA-CCI soil
moisture; and (5) the combination of Q, ET, TWS, and SM in a hindcast experiment (2004--2010).
The effects of error perturbations for both, the analysis and the forecasts are presented, and optimal trade-offs are discussed.
While large perturbations are preferred for the analysis time step, steep deterioration is observed for longer lead times, for
which more conservative error measures should be considered. From all the datasets, complementary GRACE TWS data
together with streamflow exhibit largest predictive skill for streamflow prediction itself. Stable results are obtained
with limited ensemble size of 100-200 members.