The coarse spatial resolution of global hydrological models
(typically
To this end, a 0.08
Results show that the assimilation of soil moisture observations results in the largest improvement of the model estimates of streamflow. The joint assimilation of both streamflow and downscaled soil moisture observations leads to further improvement in streamflow simulations (20 % reduction in RMSE).
Furthermore, results show that the added contribution of data assimilation, for both soil moisture and streamflow, is more pronounced when the global meteorological data are used to force the models. This is caused by the higher uncertainty and coarser resolution of the global forcing.
We conclude that it is possible to improve PCR-GLOBWB simulations forced by coarse-resolution meteorological data with assimilation of downscaled spaceborne soil moisture and streamflow observations. These improved model results are close to the ones from a local model forced with local meteorological data. These findings are important in light of the efforts that are currently made to move to global hyper-resolution modelling and can help to advance this research.
In recent decades, a number of large-scale hydrological and land-surface
models have been developed to quantify the global water cycle components, to
analyse the human influence on the global water balance, to study climate
change impact on water resources and to assess global hydrological extremes,
such as drought and flood risk. VIC (Liang et al., 1994, 1996), WaterGAP
(Alcamo et al., 2003), ORCHIDEE (d'Orgeval et al., 2008), HTESSEL (Balsamo
et al., 2009), JULES (Best et al., 2011), PCR-GLOBWB (Van Beek et al., 2011),
SURFEX-TRIP (Decharme et al., 2010, 2013) and W3RA (van Dijk, 2010; van Dijk
et al., 2014) are some examples of large-scale hydrological models recently
applied to provide water resources assessment over continental to global
domains, but their coarse spatial resolution, typically between
0.5 and 1
A possible solution to the problem may be to drive these original coarse-resolution models with high-resolution meteorological data. Several meteorological forcing data sets at a global scale are available, including the European Centre for Medium-Range Weather Forecasts – ECMWF ERA-Interim – global atmospheric reanalysis data (Dee et al., 2011), the Climatic Research Unit Time Series – CRU TS – (Mitchell and Jones, 2005), the NASA reanalysis Modern-Era Retrospective Analysis for Research and Applications – MERRA – (Rienecker et al., 2011) and the WATCH Forcing Data methodology applied to ERA-Interim reanalysis data – WFDEI – (Weedon et al., 2014). They are the result of integrating Bayesian merging of the available earth observations, in situ data sets and models to construct consistent large-scale meteorological time series. Some recent scientific efforts are conducted to improve the quality and availability of these data sets, for example increasing their spatial and temporal resolution (Cannon, 2011; Ebtehaj and Foufoula-Georgoiu, 2013; Atkinson, 2013). The use of high-spatial-resolution meteorological data would indirectly improve the resolution of the large-scale model, producing higher accuracy discharge estimates. However, when models that are designed for coarse spatial resolution are used at smaller spatial scale, issues may arise with the representation of field-scale processes. One of the major issues in this respect is the neglect of lateral flow, misleading the representation of the complex interactions between river water and groundwater (surface runoff, subsurface runoff, soil moisture state, etc.). At the moment, more research is required to understand the gain that can be obtained using these higher-spatial-resolution forcing data for uncalibrated global hydrological models at finer spatial resolutions.
Another approach to bridge the gap between the different spatial scales is to assimilate ground-based and remotely sensed observations of key water cycle variables. Higher-resolution satellite data contain information at finer spatial resolution and could be used to correct for suboptimal model performance at these finer resolutions. Multiple studies have used data assimilation techniques to obtain the best possible estimate of the hydrological system status, merging the strengths of hydrological modelling and observations and mitigating their respective weaknesses (Moradkhani, 2008; Clark et al., 2008; van Dijk et al., 2014). Among the sequential and variational data assimilation methods, the ensemble Kalman filter (Evensen, 2003) has arguably emerged as the most popular choice for assimilation into land surface and hydrological models. The various individual components of the water cycle, such as surface water (Vrugt et al., 2006; Rakovec et al., 2012), soil moisture (van Dijk et al., 2014; Wanders et al., 2014a), snow water (Sun et al., 2004; Moradkhani, 2008) and groundwater (Zaitchik et al., 2008; Tangdamrongsub et al., 2015), which influence the hydrological system in different ways, can be assimilated into the model.
Soil moisture assimilation has been considered to improve model estimates, due to its key role in the terrestrial water cycle and its responsibility for the partitioning of precipitation between surface water and storage through infiltration. Several studies have assimilated soil moisture data (Draper et al., 2011; Chen et al., 2011; Brocca et al., 2012; Wanders et al., 2014b; Massari et al., 2015; Alvarez-Garreton et al., 2015; Lievens et al., 2015) both based on ground soil moisture measurements and remotely sensed satellite soil moisture products from remote observation systems, such as ASCAT (Naeimi et al., 2009), SMOS (Kerr et al., 2012) and AMSR-E (Owe et al., 2008). On the other hand, surface water information has often been used for data assimilation frameworks (Vrugt et al., 2006; Rakovec et al., 2012) because discharge provides integrated information of all hydrological states, which is often very effective in improving model simulations. However, the risk of an integrated observation is that in some scenarios accurate simulations could be obtained by adjusting the wrong state's variables.
An additional improvement could be made by the assimilation of downscaled or disaggregated satellite soil moisture observations into a particular land surface model (Merlin et al., 2006; Sahoo et al., 2013). Recently, new soil moisture products of higher spatial resolution have been released (Gevaert et al., 2015) but their impact on hydrological model predictions has not yet been explored. For example, they could be used to correct for incorrectly observed (or interpolated) precipitation patterns, which directly affect the input uncertainty to the model.
Moreover, improved results can be obtained by assimilation of multiple observational data sets of different parts of the hydrological cycle into the hydrological model (Barrett and Renzullo, 2009; Reichle et al., 2014). For example, the joint assimilation of discharge and soil moisture could result in an improved understanding of the runoff generation mechanisms and increase the quality and quantity of information incorporated to the model system. However, the added value of this type of joint assimilation procedures is largely unknown and should be further investigated (Aubert et al., 2003; Lee et al., 2011).
Many data assimilation experiments have been set up in conjunction with local-scale hydrological models and the benefit of data assimilation for large-scale models remains largely an open question. In this context, it is interesting to analyse whether the accuracy of large-scale hydrological models can be improved and become more comparable with locally calibrated model estimates if satellite observations are assimilated. Understanding the potential gain of assimilating satellite observations into large-scale models is a relevant research opportunity and may have potential benefits for water resources management (van Dijk and Renzullo, 2011). For example, in regions without or with low-quality meteorological observations, the use of large-scale models in combination with satellite data assimilation could improve our understanding of the available water resources.The primary goal of the present study is to investigate the impact of assimilating streamflow and satellite soil moisture observations on the accuracy of global hydrological model estimations, when driven by either coarse- or high-resolution meteorological observations. The Murrumbidgee River basin in the southeast of Australia was chosen as a case study for the investigation because of the variety of land uses in the area, the high level of monitoring available for a large number of relatively unimpaired catchments and the extensive body of previous studies observing and describing the hydrologic patterns across the basin (Renzullo et al., 2014; van Dijk and Renzullo, 2011). A total of eight data assimilation scenarios were considered in which discharge and soil moisture observations were either independently or jointly used, and the forcing data were obtained from either local or global data sets. In this context, comparison of the eight scenarios with the locally calibrated model estimates provides insight into how the estimations of global hydrological models driven by global forcing data can come closer to local-scale model predictions.
Map of the Murrumbidgee River basin and its location in Australia as part of the Murray–Darling system. Green squares indicate locations for assimilation of streamflow observations and orange squares indicate locations for evaluation of streamflow observations. Each streamflow location is identified with a gauging station number according to BoM (2015). Yellow points indicate locations of field-measured soil moisture observations.
The selected study area was the Murrumbidgee River basin
(84 000 km
The simulations for two distributed hydrological models, i.e. the local OpenStreams wflow_sbm and the global PCR-GLOBWB, were performed for the period 2007–2010. The period 2000–2007 was used to spin up the models. The local and large-scale models are described in detail in the following two subsections.
The local-scale hydrological model employed in this study was the OpenStreams
wflow_sbm model (Schellekens, 2014). This is a distributed model derived
from the topog
OpenStreams wflow_sbm model structure, adapted from Vertessy and
Elsenbeer (1999) and Schellekens (2014). Symbols' definitions are as follows:
Precip, precipitation; Evap, evaporation;
OSWS model structure consists of three main routines: (i) rainfall interception (schematized by the Gash model – Gash, 1979), (ii) soil processes (schematized by the topog_sbm model) and (iii) river drainage and overland flow (modelled by the kinematic wave routing over a drainage network).
The water enters each model cell from precipitation to the canopy interception storage or snow storage. The remaining liquid water infiltrates the soil. At the same time, water is taken from the soil through evapotranspiration (based on soil water content and vegetation type). The water exchange into the soil considers two vertical soil layers, the unsaturated zone (UZ) and the saturated zone (SZ), based on topog_sbm structure. Total runoff is the sum of the direct runoff, the melt water that does not infiltrate into the soil and the baseflow (lateral subsurface flow from the saturated zone). This total runoff is conducted along the river network as discharge with kinematic wave routing.
The OSWS model was calibrated for the Murrumbidgee River basin using observations from in situ streamflow gauges (BoM, 2015) for the time period 1990–2010. These gauges are different from those considered in all the data assimilation scenarios to ensure an independent verification.
The large-scale hydrological model employed in this study was PCR-GLOBWB (Van
Beek and Bierkens, 2009; Van Beek et al., 2011). Similar to OpenStreams
wflow_sbm, PCR-GLOBWB is essentially a leaky-bucket type of model applied
on a cell-by-cell basis. PCR-GLOBWB is coded in the PCRaster-Python
environment. A spatial resolution of 0.08
PCR-GLOBWB model structure, adapted from Van Beek et al. (2011).
Symbols' definitions are as follows: Precip, precipitation; Evap, evaporation;
For each time step and cell, PCR-GLOBWB calculates the water balance components, including the water storage in three vertical soil layers (0–5, 5–30 and 30–150 cm) and one underlying groundwater reservoir, as well as the water exchange between the layers (percolation, capillary rise) and between the top layer and the atmosphere (rainfall, evapotranspiration and snowmelt). Subgrid variability is taken into account considering the variations of elevation, land cover, vegetation and soil. The total runoff of a cell consists of direct runoff (saturation excess surface runoff), non-infiltrating melt water, interflow (lateral drainage from the soil profile) and baseflow (groundwater runoff from the lowest linear reservoir). The simulated runoff is routed along the river network based on the Simulated Topological Networks (STN30; Vörösmarty et al., 2000). Water abstraction and consumptive water use (domestic, industrial, livestock, irrigation) and reservoir management are included.
Daily mean precipitation, temperature and reference
evapotranspiration for the time period 2007–2010 from the (i) global
(0.5
In contrast to the local-scale model, PCR-GLOBWB was not calibrated for the study basin. Hydrological model parameters were derived from vegetation, soil properties or geological information and estimated at a global scale.
The forcing data required to drive both hydrological models are precipitation and air temperature. Two types of forcing data were used in this study: local forcing data, representing the best available data, and global forcing data representing a lower spatial resolution data set but one which is available globally.
Local precipitation and air temperature data were obtained from the gridded
data sets generated by the Australian Bureau of Meteorology under the
Australian Water Availability Project (AWAP) (Jones et al., 2009). The data
are derived from station-level daily precipitation and air temperature, and
interpolated on a 0.05
Global precipitation and air temperature data were obtained from the WATCH
Forcing Data methodology applied to ERA-Interim reanalysis data – WFDEI – (Weedon
et al., 2014). The daily global precipitation and air temperature
data were provided at a spatial resolution of 0.5
Local and global reference evapotranspiration (ET) were obtained through Hamon method (Allen et al., 1998; Lu et al., 2005).
Figure 4 shows the daily mean precipitation, temperature and reference evapotranspiration for the study time period (2007–2010) for both forcing data sets. Aside from the resolution difference, precipitation ranges from higher values in the mountainous regions of the catchment, with increasing variance in elevation and rainfall, to lower values in the western plain. On the contrary, temperature and, subsequently, reference evapotranspiration decrease from west to east.
Catchment daily mean (
Local and global forcing data show some differences in their spatial
distribution and magnitude. Each climate variable shows similar spatial
distribution across the various resolutions (rows in Fig. 4), with larger
variations in the high elevation areas. However, local temperature magnitude
differs by 3–4
Table 1 shows the catchment daily mean values of the climate forcing
variables for each year individually. Global temperature deviates from local
by approximately 3–4
Soil moisture observations retrieved from AMSR-E (Advanced Microwave Scanning
Radiometer – EOS) brightness temperatures were provided by the Vrije
Universiteit Amsterdam (VUA) in collaboration with NASA. AMSR-E is
a multi-frequency passive microwave radiometer that uses C- (6.9 GHz)
and X-band (10.65 and 18.7 GHz) radiance observations to derive
near-surface soil moisture via the LPRM radiative transfer model (Owe et al.,
2008; De Jeu et al., 2008). In the present study, C-band AMSR-E data reported
on a regular
Soil moisture observations from AMSR-E were compared to the unsaturated zone
layer of OSWS and the first of the three vertical layers constituting the
soil profile in each grid cell of PCR-GLOBWB. To match the remotely sensed
soil moisture observations to the statistics of corresponding hydrological
model states for soil water, different strategies can be followed, such as
linear or minimum–maximum (MM) matching (Brocca et al., 2011), mean–standard
deviation (
Soil moisture monitoring sites' information.
In situ soil moisture observations were obtained from the Australian moisture
monitoring network, OzNet (
Hydrometeorological and geographical information of analysed catchments at the Murrumbidgee River basin.
Discharge observations were provided by the Bureau of Meteorology (BoM) and
the Commonwealth Scientific and Industrial Research Organisation (CSIRO),
under the Water Information R
The ensemble Kalman filter (EnKF) is a sequential data assimilation method
evolved from the standard Kalman filter (Evensen, 1994) that has been used
previously for assimilation of observations into land surface and
hydrological models (Chen et al., 2011; Draper et al., 2011; Wanders et al.,
2014a; Tangdamrongsub et al., 2015). It is a Monte Carlo-based approach that
integrates an ensemble of model states forward in time to represent the error
statistics of the model estimate when observations are assimilated (Burgers
et al., 1998; Evensen, 2003). The state equation in a discrete form is given as
Data assimilation scenarios including abbreviations, forcing data, hydrological model and assimilated observations.
The EnKF calculates the analysis at each time
In this study, the EnKF was applied to update state variables of the
large-scale hydrological model, PCR-GLOBWB, on each daily time step using
downscaled remotely sensed AMSR-E soil moisture and discharge observations.
We used 100 ensemble members and all the observations were calculated and
assimilated at each 0.08
A total of eight different data assimilation scenarios with PCR-GLOBWB were
intercompared and compared to the OSWS estimates without any data
assimilation. The data assimilation (DA) scenarios are described in Table 4,
indicating the meteorological forcing and the observations used in each
scenario. Simulations forced with local meteorological data are denoted with
LOCAL and simulations forced with global meteorological data are denoted with
GLOBAL. Independent assimilation of discharge (GLOBWB_Q) and soil moisture (GLOBWB_SM)
were investigated, as well as the joint assimilation of both observation types (GLOBWB_SM
In the EnKF, to account for model and observations uncertainty, stochastic
noise can be introduced in model forcing data, parameters, soil moisture and
discharge observations. For the local and global meteorological forcing, the
precipitation was perturbed with additive Gaussian white noise with standard
deviation of 10 % of the nominal value (Adam and Lettenmaier, 2003;
Hijmans et al., 2005). The errors were assumed to be spatially uncorrelated.
For the assimilation of the satellite soil moisture data, spatial information
on the measurement error covariance
Simulated and observed soil moisture estimates at gauging
station no. 410057 in a tributary of the Murrumbidgee River for the time period
January 2008–May 2009. The upper panel shows soil moisture time series when
local data are used as model forcing. Soil moisture time series obtained with
the global forced models are shown in the lower panel. Each panel contains
results for each data assimilation scenario plotted with different coloured
lines (OSWS – orange, GLOBWB_Q – blue, GLOBWB_SM – green and
GLOBWB_SM
The impact of assimilating discharge and soil moisture observations into the
large-scale hydrological model PCR-GLOBWB, compared with the locally
calibrated model OSWS, is separately analysed: firstly, on soil moisture
estimates and secondly, on discharge estimates. A common regular
0.08
Results were produced for each of the 23 locations listed in Table 3. For practical reasons, the following section includes results for a limited number of evaluation locations only, both in the Murrumbidgee River and its tributaries. This combination thus comprises stations with varying sizes of contributing area.
To understand and intercompare the performance of the different data
assimilation scenarios described in Table 3, an extensive evaluation was
carried out, including the calculation of various evaluation metrics, such as
root mean squared error (RMSE), mean absolute error (MAE), Pearson
correlation coefficient (
The time series of simulated soil moisture for 0–5 cm for each data assimilation scenario and downscaled AMSR-E observations for the time period January 2008–May 2009 at gauging station no. 410057 (mountainous region) are shown in Fig. 5.
The use of global forcing data produces a different dynamic response of soil moisture estimates compared to the local forcing data. This fact is due to the discrepancies between both meteorological data sets, which govern the water exchange processes between the top layer and the atmosphere (precipitation, evapotranspiration and temperature).
Even though global precipitation quite accurately depicts the overall character of the precipitation (daily mean values of the local and global precipitations show similar spatial distributions and magnitudes; see Sect. 2.3.1), the global precipitation misses specific rainfall events at particular days and locations due to its lower resolution. This is especially important for warm season precipitation and regions in mountainous terrain (e.g. gauging station no. 410057 in Fig. 5), which are dominated by convective storms. The differences in precipitation are reflected in soil moisture estimates of both hydrological models and their impact is higher in PCR-GLOBWB estimates. When PCR-GLOBWB is forced with local data, soil moisture estimates produce patterns with a more accurate description of the small-scale variability of the observations in time. Whereas when global forcing is used, soil moisture results in a smoother estimation of the observations.
Evaluation results of the catchment daily means of soil moisture in
the Murrumbidgee River basin. In the rows, three different evaluation metrics
are shown; from top to bottom, these are the MAE, the RMSE and
A total of eight different data assimilation scenarios were investigated, with global and local forcing data (Table 3). In the OL scenario, no data are assimilated into the model to correct intermediate model states. In this first scenario, soil moisture observations are underestimated when local forcing is used. The assimilation of discharge observations, Q scenario, results in similar soil moisture estimates compared to the OL scenario, as expected. This similarity is caused by the assimilation procedure which constrains the model to follow the discharge observations directly affecting groundwater and routing processes, which are a poor reflection of the upper soil moisture content from 0 to 5 cm. In contrast, the assimilation of soil moisture observations in SM and SM_Q scenarios produces a reduction of the negative bias, improving the soil moisture estimates, especially when local forcing is used.
Figure 6 shows the impact of each data assimilation scenario on the
considered evaluation metrics (RMSE, MAE and
From this figure, some general observations can be made. Evaluation results
show differences between the results from local and global forcing of the
models. For example, the use of local instead of global forcing produces
a decrease of 4 and 2 % in RMSE and MAE, respectively, and an increase of
6 % in
Boxplots of the catchment daily means of soil moisture in the Murrumbidgee River basin. The upper panel shows soil moisture when local data are used as model forcing. Soil moisture obtained with the global forced models is shown in the lower panel. Boxplots of each panel illustrate the first and third quantile ranges (box), the median (dark line) and the maximum–minimum range (whiskers) of soil moisture estimates.
Simulated and observed streamflow estimates at gauging
station no. 410088 in a tributary of the Murrumbidgee River for the time period
January 2009–January 2010. The upper panel shows streamflow when local data
are used as model forcing. Streamflow obtained with the global forced models
is shown in the lower panel. Each panel contains results for each data
assimilation scenario and the observed streamflow estimates plotted with
different coloured lines (OSWS – orange, GLOBWB_OL – red, GLOBWB_Q – blue,
GLOBWB_SM – green, GLOBWB_SM
The large-scale model, PCR-GLOBWB, without data assimilation shows a poorer
performance than the locally calibrated model OSWS on soil moisture
predictions. Assimilating discharge observations (Q scenario) does not lead
to an improvement of soil moisture estimates, whereas downscaled AMSR-E soil
moisture observations assimilation (SM scenario) results in an increase of
Additionally, boxplots of the catchment daily mean values are included in Fig. 7 considering local (upper panel) and global (lower panel) forcing. The assimilation of soil moisture observations leads to an evident improvement in the statistical distribution of PCR-GLOBWB soil moisture estimates, reducing the differences in dispersion with the observations.
Assimilating soil moisture observations and forcing the model with high-spatial-resolution
meteorological data sets impacts the quality of soil
moisture estimates with PCR-GLOBWB to a similar extent. Results indicate that
the highest improved performance is achieved when their combination occurs,
i.e. soil moisture observations are assimilated into a model driven by local
forcing data (LOCAL GLOBWB_SM and LOCAL GLOBWB_SM
The simulated and observed streamflow estimates at gauging station no. 410088 are shown in Fig. 8. From this figure, it is clear that the peaks in streamflow are poorly estimated by PCR-GLOBWB, whereas OSWS is able to capture them with higher accuracy, independently of the forcing data used. This is most probably explained by the higher resolution and the calibration of model parameters for the study.
Differences between local and global forcing data (see Sect. 2.3.1) are reflected in differences in streamflow estimates from both models. When global data are used, evapotranspiration is lower; hence, a higher amount of water is introduced into the models, resulting in higher streamflow estimates. By assimilating discharge and soil moisture observations, intermediate hydrological processes, including groundwater state, percolation and surface runoff, among others, are corrected and errors in forcing data are reduced to improve streamflow predictions (e.g. November 2008). The best performance is achieved when soil moisture and discharge data are jointly assimilated.
To further analyse and quantify the influence of each data assimilation
scenario on streamflow estimates, the evaluation metrics (RMSE, MAE,
The highest
Both observations' assimilation, discharge and soil moisture separately (Q and
SM), improve streamflow models' estimates. The highest improvement is achieved
when both observations are assimilated into the model (GLOBWB SM
Evaluation results for streamflow estimates at station nos. 410001, 410005,
410078 and 410136 in the Murrumbidgee River. Average values
calculated across those locations are shown in the rightmost bar of each
histogram. In the rows, four different evaluation metrics are shown; from top
to bottom, these are the MAE, RMSE,
The largest improvements were found at gauging stations in the main channel
of the Murrumbidgee River, such as station no. 410001 where assimilating soil
moisture and discharge observations increases
Boxplots of streamflow estimates are included in Fig. 10. The results clearly
show that, compared to the observed streamflow, the median values of
PCR-GLOBWB streamflow without data assimilation are very biased. The greatest
amount of spread is observed between OL scenarios and observations, across
all stations. Every data assimilation scenario (Q, SM and SM
In contrast with the soil moisture evaluation results, Figs. 8–10 indicate that for streamflow the use of local forcing provides a larger improvement than assimilating soil moisture and discharge observations. Using a global model with local forcing and assimilating satellite soil moisture data yields streamflow predictions comparable to a local model with local forcing along the main river of this catchment. Moreover, also on the main channel, a global model with global forcing may still yield reasonable results as long as both discharge data and soil moisture data are assimilated.
Boxplots of streamflow estimates at station nos. 410001, 410005, 410078 and 410136 in the Murrumbidgee River. The upper panel shows streamflow when local data are used as model forcing. Streamflow obtained with the global forced models is shown in the lower panel. Boxplots of each panel illustrate the first and third quantile ranges (box), the median (dark line) and the maximum–minimum range (whiskers) of streamflow estimates.
PCR-GLOBWB poorly estimates streamflow and soil moisture, when forced either with high- or coarse-spatial-resolution forcing data, without data assimilation. The derivation of the hydrological model parameters from hydro-geological information at a global scale could be a possible explanation. From the initial scenarios without assimilation, it can be inferred that there is significant space for improvements when discharge and soil moisture observations are assimilated. An alternative route to data assimilation to improve model estimates would be to locally calibrate PCR-GLOBWB using discharge observations from in situ gauging stations. The improvement achieved through model parameters calibration would be possibly even higher than when soil moisture and streamflow observations are assimilated into the model (Wanders et al., 2014b). However, the present study aims to be an attempt at providing hydrological estimations with a global model that could be also used in ungauged river basins where scarce in situ data are available.
The joint assimilation of discharge and downscaled satellite soil moisture observations produces the largest improvement on PCR-GLOBWB streamflow estimates (20 % reduction in RMSE). These results agree with the findings made by Wanders et al. (2014a) in the Upper Danube and are also in line with the expectations where more observations lead to a better constrained model simulation.
A major finding of this study is the positive impact of assimilating soil moisture observations on the streamflow estimates, compared to the independent discharge assimilation. The estimations of a large-scale hydrological model driven with coarse-resolution forcing data are improved using globally available remotely sensed soil moisture observations. This creates an opportunity for improved global model-based streamflow estimations as a result of the assimilation of remotely sensed soil moisture observations into large-scale hydrological models. Additionally, the adopted methodology has the advantage that it is applicable in river basins all over the world independently of the availability of in situ hydrometeorological information.
The variable effectiveness of soil moisture assimilation has been previously reported in literature. Whereas some studies found improvements (Draper et al., 2011; Wanders et al., 2014b), others obtained mixed or unsatisfactory results (Crow et al., 2005). The scale of soil moisture observations, the dominant runoff processes in the study basin and the model structure and parameter uncertainties may partly explain this variability. In this particular study, the novel use of a finer-spatial-resolution satellite soil moisture product together with the climate and hydrogeological characteristics of the catchment could be a possible explanation of the positive impact of soil moisture assimilation. Renzullo et al. (2014) assimilated satellite soil moisture observations from multiple sensors (ASCAT and AMSR-E) obtained at scales coarser than the model (AWRA-L) resolution. In the present study, the scale of soil moisture observations coincides with the model scale. A specific analysis of the impact on streamflow and soil moisture estimates of assimilating non-downscaled AMSR-E soil moisture assimilation could be a possible route to further investigate the effect of different spatial-resolution soil moisture products. To this end, AMSR-E soil moisture observations at the original spatial resolution were assimilated into PCR-GLOBWB and results analysis presented in this study were reproduced. Results showed that the assimilation of non-downscaled soil moisture observations has a positive impact on soil moisture and discharge estimates, but this improvement is smaller compared to the assimilation of the downscaled soil moisture estimates (Figs. S7 and S8). Moreover, runoff in the Murrumbidgee River basin is mainly dominated by direct runoff processes, with reduced contribution from the groundwater zone (Green et al., 2011). These catchment conditions, together with their representation in the model structure are most likely responsible for the added value of assimilating soil moisture. There may be merit in analysing these scenarios in future research studies.
For the assimilation of AMSR-E soil moisture observations, spatial
information on the measurement error covariance
Meteorological data play a key role in soil moisture and discharge model estimates. The various model and data assimilation options were evaluated under both high- and low-resolution meteorological forcing. In general, the higher spatial resolution and the higher quality of the local forcing data results in better model predictions (PCR-GLOBWB and OSWS) of both soil moisture and streamflow. The coarse resolution of the global forcing could lead to failures to detect extreme rainfall events or differences at specific regions, such as mountainous areas. Recent studies have developed several downscaling procedures, e.g. geostatistical methods of blending satellite and gauge data (Chappell et al., 2013), that could improve model predictions when global forcing is used.
To improve the representation of the global water cycle using global hydrological models, one could follow multiple strategies: improve the quality and quantity of ground observation, increase the spatial resolution of the global models or obtain more detailed information on the catchment properties (e.g. soil data). Another way forward is the assimilation of observations and the use of high-spatial-resolution meteorological data to bridge the gap between the different spatial scales, for which large-scale hydrological models are designed, and the river basin scale. The advantage of this approach is that it provides a global improvement of the hydrological simulation and the satellite data often have a global coverage. In this study, we show the potential gain in model accuracy of using remotely sensed observations. Previous studies on the potential gain of satellite observations for global and continental hydrological models agree with the obtained results in the present study (Andreadis and Lettenmaier, 2006; Lievens et al., 2015).
The study investigates the influence of discharge and soil moisture assimilation on the accuracy of large-scale model (PCR-GLOBWB) predictions, when driven by meteorological forcing data sets of high and coarse resolution, compared with local-scale model (OSWS) estimations.
Results show poor PCR-GLOBWB streamflow and soil moisture estimates when no observations are assimilated. The assimilation of soil moisture observations results in the largest improvement of the model estimates of streamflow. The joint assimilation of both streamflow and downscaled soil moisture observations leads to further improvement in streamflow simulations (20 % reduction in RMSE).
In general, the higher spatial resolution of the local forcing data results in better models predictions of both soil moisture and streamflow. The added value of using higher-spatial-resolution forcing data is more significant for PCR-GLOBWB than for OSWS. When the impact on model accuracy of assimilating observations and forcing the models with higher-spatial-resolution data are compared, the latter leads to a more substantial improvement of streamflow predictions.
Furthermore, results show that the added contribution of data assimilation, for both soil moisture and streamflow, is more pronounced when the global meteorological data are used to force the models. This is caused by the higher uncertainty and coarser resolution of the global forcing.
The greatest benefit is obtained when local coarse-resolution forcing data are used in combination with streamflow and soil moisture observations' assimilation into the large-scale hydrological model, PCR-GLOBWB.
In conclusion, the present research study shows that data assimilation of high-resolution soil moisture succeeds in resolving shortcomings that exist nowadays in global hydrological models and can partly overcome the difference in model performance between a large-scale hydrological model driven by coarse-resolution forcing data and a local-scale model forced with higher-resolution meteorological data. Moreover, it demonstrates that further investments and improvements in remotely sensed observations, especially in soil moisture products can benefit large-scale hydrological model predictions and bring these closer to those obtained from local-scale hydrological modelling.
Evaluation results of the catchment daily means of soil moisture estimates (the exact values of Fig. 6 are included in this table for clarification).
This research received funding from the European Union Seventh Framework Programme (FP7/2007-2013) under grant agreement no. 603608, Global Earth Observation for integrated water resource assessment: eartH2Observe. The work of N. Wanders was supported by NWO (GO-AO/30) and Rubicon grant (825.15.003). We would like to thank CSIRO in Canberra, ACT, Australia for their active collaboration and constant engagement in the present research project. The authors also want to thank Anouk Gevaert for providing the downscaled AMSR-E soil moisture observations. Geert Sterk is acknowledged for productive comments and discussions on this investigation. Edited by: G. Jewitt Reviewed by: two anonymous referees