ESA CCI Soil Moisture Assimilation in SWAT for Improved Hydrological Simulation in Upper Huai River Basin

*e assimilation of satellite soil moisture (SM) products with coarse resolution is promising in improving rainfall-runoff modeling, but it is largely impacted by the data assimilation (DA) strategy. *is study performs the assimilation of a satellite soil moisture product from the European Space Agency (ESA) Climate Change Initiative (CCI) in a physically based semidistributed hydrological model (SWAT) in the upper Huai River basin in China, with the objective to improve its rainfall-runoff simulation. In this assimilation, the ensemble Kalman filter (EnKF) is adopted with full consideration of the model and observation error, the rescaling technique for satellite SM, and the regional applicability of the hydrological model. *e results show that the ESA CCI SM assimilation generally improves the streamflow simulation of the study catchment. It is more effective for low-flow simulation, while for very high-flow/large-flood modeling, the DA performance shows uncertainty. *e less-effective performance on largeflood simulation lies in the relatively low dependence of rainfall-runoff generation on the antecedent SM as during which the SM is nearly saturated and the runoff is largely dominated by precipitation. Besides, the efficiency of DA is deteriorated by the dense forest coverage and the complex topography conditions of the basin. Overall, the ESA CCI SM assimilation improves the streamflow simulation of the SWATmodel in particular for low flow.*is study provides an encouragement for the application of the ESA CCI SM in water management, especially over low-flow periods.

One promising approach to improving SM estimation in turn improving rainfall-runoff modeling is to integrate the observed SM into the hydrological modeling process using data assimilation (DA) techniques [11][12][13][14][15][16].In general, the SM data for integration can be obtained from field measurements and satellite observations.e in situ measurements are insufficient in the availability and spatial representativeness due to the high spatial heterogeneity of SM.Major researches on in situ SM assimilation focus on discussing the DA approaches and exploring the potential of SM assimilation in improving the hydrological process [13,17,18].However, the satellite observations are capable of capturing the spatial distribution and temporal dynamics of SM on large scales.Despite the fact that the satellite remote sensing (RS) can only detect the surface SM information with a few centimeters (∼5 cm), it could represent the fastest response of SM dynamics to meteorological conditions [19].A large number of studies have been implemented to assimilate the RS SM in the land surface model for the purpose of obtaining a more accurate and reliable profile SM data set on a regional or global scale [20][21][22][23][24][25][26].Nevertheless, the assimilation of coarse-scale RS SM in the hydrological model targeted at improving the rainfall-runoff process is implemented in relatively few studies [10,[27][28][29][30][31].Currently, there is still no consensus on the improvement of streamflow modeling through satellite soil moisture assimilation [4,7].For instance, almost no improvement of stream ow simulation was obtained by Brocca et al. [4] in the assimilation of the surface ASCAT SM retrievals, while up to 10-30% improvements were achieved in such other studies as Massari et al. [32], Lopez et al. [7], and Loizu et al. [10].
e large discrepancies of the DA performance in previous studies are likely due to the fact that it is in uenced by various factors in the DA framework setup, such as the quanti cation of the model and observation error, the mismatch between the observed and simulated SM, the data quality and rescaling technique for RS SM, and the model physical mechanism and its regional applicability.To date, the added value of satellite soil moisture data in hydrological modeling is still underexplored [5,32].
e performance of RS soil moisture assimilation in stream ow modeling presents certain speci city on the satellite data itself, the hydrological model, and the di erent con guration schemes in the DA framework setup.erefore, speci c studies on satellite soil moisture assimilation with comprehensive consideration of the DA implementation strategies are essential for exploring the signi cance of satellite soil moisture in hydrological modeling.
In this paper, a case study for satellite soil moisture assimilation is implemented in the upper Huai River basin in China, with full consideration of the factors in the DA framework including the quanti cation of the model and observation error, the rescaling technique for RS SM, and the regional applicability of the hydrological model.is data assimilation is performed in a physically based semidistributed hydrological model (SWAT) based on a robust sequential data assimilation approach (the ensemble Kalman lter (EnKF)).A multisatellite-merged soil moisture data set from the European Space Agency (ESA) Climate Change Initiative (CCI) is adopted as the assimilation data source.2 Advances in Meteorology e main objective of this study is to explore the potential of coarse-scale RS soil moisture in improving runo modeling and to provide recommendations on the assimilation strategy.

Study Area.
e study catchment is located in the upstream basin of the Huaibin hydrologic station in the Huai River basin, China (Figure 1).e watershed covers about 16000 km 2 .e whole watershed is located in the transition region between the northern subtropical zone and the warm temperate zone.Its annual average rainfall is around 900 mm, 50%-80% of which falls during June to September.Here, the annual average temperature is about 15 °C.e major land cover is the agriculture land (AGRC 32.5%, RICE 35%) and forest (FRST 23.6%) (Figure 2).

Data for SWAT Model.
SWAT model building requires meteorological and underlying surface data.e meteorological data mainly include precipitation, maximum and minimum temperature, solar radiation, wind speed, and relative humidity.e precipitation data are drawn from 106 local rainfall stations within the catchment (Figure 1).e other ve meteorological data come from the three meteorological gauges (Xinyang, Gushi, and Guangshui) (Figure 1).e underlying surface data are the digital elevation (DEM), land cover, and soil category data.e DEM data are downloaded from the Shuttle Radar Topography Mission with a spatial resolution of 90 m (http://datamirror.csdb.cn/index.jsp).
e land use/land cover (LU/LC) data are collected from the Chinese Cold and Arid Regions Science Data Center (http://westdc.westgis.ac.cn/) with a spatial resolution of 1 km (Figure 2).e soil data are resampled from a soil map at a scale of 1 : 100000 from the Soil Handbook of Henan Province.
e soil for the whole catchment is divided into seven categories (Figure 2).e soil texture and its corresponding United States Department of Agriculture (USDA) classi cation for each category are shown in Table 1.Besides, there are six hydrologic stations (Dapoling, Changtaiguan, Zhuganfu, Xixian, Huangchuan, and Huaibin) (Figure 1) with daily stream ow measurements of 1992-2008 (the data quality issue exists for the years 2000 and 2001) in this basin.

ESA CCI Soil Moisture Data.
e ESA CCI soil moisture data are a merged multisatellite surface soil moisture product developed in the Climate Change Initiative (CCI) by the European Space Agency (ESA).It combines the soil moisture retrievals from four microwave radiometers (SMMR, SSM/I, TMI, and AMSR-E) and two scatterometers (AMI and ASCAT) into a 0.25 °global daily data set over 30 years from 1978.
e data integration relies on their respective sensitivity to vegetation density and uses a Noah GLDAS-1 surface soil moisture product [33] as a climatology reference [34].e ESA CCI SM consists of active, passive, or combined products.e active/passive products are the integration of the scatterometer/radiometer-based SM retrievals, respectively, while the combined product is the fusion of both the active and passive products.In this study, the combined product (ESA CCI SM v03.2) is adopted for soil moisture assimilation.

Soil and Water Assessment Tool (SWAT).
e SWAT is a physically based semidistributed watershed model, which has been widely used in rainfall-runo modeling over recent years [35,36].In hydrological modeling, the catchment is rstly delineated into several subbasins according to its topography.en, each subbasin is further divided into several hydrological response units (HRUs) based on the land use, soil, and slope.HRUs are basic calculation units for the land phase of the hydrologic cycle, on which the processes for surface runo , lateral ow, and ground water are generated accompanied by evapotranspiration and soil water routing.
Soil moisture lies in the center of the hydrologic cycle and makes di erent impacts on the above process.e initial pro le soil water content in uences surface runo generation through the curve number in the SCS method [37].After surface runo generation, the water in ltrated to the

Advances in Meteorology
soil profile is redistributed based on a storage routing technique with the soil water field capacity as the threshold.e water balance for each soil layer can be expressed as follows: where SW ly and SW ly ′ are the soil water content (mm) at the start and end of the day, Δw perc,ly is the net percolation from the overlying layer (i.e., the layer ly + 1), Q lat,ly is the lateral flow generated from the layer ly, and E a,ly is the evapotranspiration drawn from the layer ly. e evapotranspiration from soil mainly includes two parts: soil evaporation and plant uptake/transpiration.In the SWAT, the potential evapotranspiration is firstly calculated using the Penman-Monteith equation [38].Based on the potential evapotranspiration, the leaf area index, and the aboveground biomass and residue conditions, both the demand for transpiration/plant uptake and the demand for soil evaporation are determined.en, the soil evaporation demand and the plant uptake demand for each soil layer are estimated using a depth distribution function.Finally, relying on the soil evaporation demand and the plant uptake demand with the available soil water as a constraint, the actual soil water evaporation and plant uptake are determined.In the processes mentioned above, the actual soil water extraction of a given layer is not allowed to be compensated by the extraction from other layers.However, the soil water deficiency can be made up by adjusting the soil compensation (esco) and plant compensation (epco) factors via changing the depth distribution of the soil evaporation demand and the plant water uptake demand.Besides, the calculation for the soil water percolation (w perc,ly ) and lateral flow (Q lat,ly ) is [39] omitted here.
In the water routing phase, the SWAT adopts a storage feature to calculate the surface runoff and lateral flow generated from each HRU to the main channel, while it applies a linear reservoir similar technique to account for the ground water to the main channel.
e channel water routing is performed using a variable storage routing method [40].

e Ensemble Kalman Filter (EnKF) for Soil Moisture
Assimilation.
e EnKF is a sequential DA approach evolved from the standard Kalman filter [41].It is based on an ensemble of model states produced by adding the Monte Carlo noise to model forcing and states and/or parameters to approximate the model state error covariance matrix for the purpose of optimally merging the model predictions with observations.e state ensemble forecast at time t can be expressed as follows: where X f t is the forecasted state ensemble at time t and X u t−1 is the updated state ensemble at t − 1.In this study, it is constructed by the profile SM with up to four layers (Table 1) for all HRUs of the study basin (the HRUs delineation is detailed in Section 3.1).u t represents the model forcing inputs.In this study, it mainly includes the observed precipitation P and temperature T at each site.e precipitation error is assumed to be independent both in time and in space; that is, both the autocorrelation between time steps at each rainfall station and the error correlation among different stations are ignored.e perturbation (η p ) to precipitation is assumed to be a lognormal multiplicative distribution with mean 1 and covariance σ 2 p (3). e perturbation to temperature (η T ) is assumed to be an additive normal distribution with mean 0 and covariance σ 2 T (4).Besides, δ represents the model parameter with a perturbation (η par ) of normal multiplicative distribution of mean 1 and covariance σ 2 par (5).w t is the stochastic perturbation to the forecasted SM, being assumed to be an additive normal distribution with mean 0 and covariance σ 2 s : e state update for soil moisture can be obtained by where Z t is the observation ensemble at time t.It is constructed by the RS SM for all grids covering the basin and being stochastically perturbed by an additive normal distribution with mean 0 and covariance σ 2 R .H is the observation operator, being used to map the model states to the observations.It is constructed by the area proportions of HRUs in RS grids as the SWAT model-simulated SM is on the HRU level, while the observed SM is on RS grids.K t is the Kalman gain, which is calculated based on the forecast and observation error covariance: where P ms,t is the cross-error covariance between the predicted SM (X f t ) and the measurement prediction H(X f t ) at time t,  3.

SWAT Calibration.
e SWAT model for the upper Huai River basin is built up based on the meteorological forcing and land surface data (Section 2.2). is catchment is partitioned into 37 subbasins (Figure 1) and 146 hydrological response units (HRUs).To simplify and improve the model calibration process, a sensitivity analysis of the model parameters to the hydrologic modeling is implemented using the Latin-hypercube and one-factor-at-a-time method [42,43].ereafter, the model is calibrated and validated using the daily runo records over 1992-1999 and 2002-2008 at the interior and outlet hydrologic stations (i.e., Dapoling, Changtaiguan, Zhuganfu, Xixian, Huangchuan, and Huaibin in Figure 1), respectively.e parameter optimization is achieved by a combination of the autocalibration using the Sequential Uncertainty Fitting (SUFI2) [44] with the Nash-Sutcli e coe cient of e ciency (detailed in Section 3.4) as the objective criteria and the manual ne-running method.

Model Error Estimation.
e determination of the model error is signi cant for the performance of DA as the model predictions and observations are merged based on the relative weight between the model and observation error (as in ( 6) and ( 7)).In this study, the model error is mainly contributed by the model input error for precipitation and temperature, the model parameter error for parameters sensitive to SM simulation (e.g., the available soil water capacity), and the model state error for simulated SM. e various errors mentioned above are characterized by additive/multiplicative normal/lognormal distribution speci ed in Section 3.2.e assumed distribution is only controlled by the standard deviation (SD), that is, σ p , σ T , σ par , and σ s in (3), ( 4), (5), and (2), respectively.
erefore, the model error estimation is to quantify the model error parameters σ p , σ T , σ par , and σ s .
e quanti cation of σ p , σ T , σ par , and σ s is performed by analyzing the statistical characteristics of the simulated stream ow ensemble driven by model error perturbations on the basis of the observed stream ow at in situ sites.If the ensemble spread after perturbation is too large, the overtting of observation exists in DA.Otherwise, the observed information cannot be fully utilized in DA. erefore, the two ensemble veri cation measures ( 8) and ( 9) should be satis ed [32].at is, if the ensemble spread sp is large enough, the temporal mean of the ensemble skill sk should be similar to the temporal average of the ensemble spread sp: And the observation should be indistinguishable from a member of the ensemble (N is the ensemble size): where where Q i k is the model simulated stream ow of the ensemble member i at time k, Q k is the ensemble mean of the model simulated stream ow, Q obs,k is the observed stream ow at time k, and T is the total time step.Di erent sp/sk and 〈sk〉/〈mse〉 can be obtained with di erent σ p , σ T, σ par , and σ s , that is, f(σ p , σ T , σ par , σ s ) sp/sk and g(σ p , σ T , σ par , σ s ) 〈sk〉/〈mse〉.erefore, the optimal estimation of σ p , σ T , σ par , and σ s can be realized by searching for the minimum value of the following function: In

Advances in Meteorology
In the soil moisture assimilation, the temperature error (σ T ) is not very sensitive to DA performance.Hence, σ T is set to be 1 °C as referenced from the study of Chen et al. [12].Besides, the predicted/simulated SM error σ s is set to be 0.01 m 3 /m 3 to avoid rapid changes of soil water content between continuous time steps [45,46].Finally, the precipitation (σ p ) and model parameter error (σ par ) are determined by searching for the minimum F based on the streamflow measurements at the catchment outlet (Huaibin station) over 2002-2004 with an ensemble size of 200.

ESA CCI SM Bias Correction.
Remote sensing (RS) retrieval of SM often has systematic bias to the in situ observed and model-simulated soil moisture due to their large differences in spatial resolution and detection depth.e model-simulated SM can generally meet the water balance of the basin/region.In order to keep the basin's water balance in DA, the systematic bias in RS SM needs to be corrected before DA [47].In this study, the bias correction for the ESA CCI SM uses the cumulative distribution function (CDF) approach [48], where the probability of the RS SM and the simulated SM is assumed to be the same.
e spatial matching between them uses the area-weighted average method to aggregate the simulated SM from HRUs to RS grids.Here, the rescaling is performed over the complete model validation period of 2002-2008, considering that CDF estimation typically requires a long record of observed and model simulated data [6].

ESA CCI SM Error Estimation.
Rational quantification on the uncertainty of RS SM is important for its optimal application.RS SM still has considerable uncertainty although its accuracy and reliability have been largely improved in recent years [49,50].In this study, referencing from previous researches [5,51,52], the error for ESA CCI SM is assumed to be an additive Gaussian distribution with the standard deviation (SD) of σ R .Here, the estimation of σ R is obtained from the equation referring to the study of Lievens et al. [30]: where sm_uncertainty is an indicator of the data uncertainty for the ESA CCI SM [53,54], which is not fully considered as the representativeness error (e.g., the error caused by vegetation or different layer depths).e representativeness error is accounted by the parameter a 0 , which represents the minimum retrieval error for ESA CCI SM. frc is the fraction of the ESA CCI SM grid cell covered by the forest.e calculation of frc is based on the land cover data collected from the Chinese Cold and Arid Regions Science Data Center (http://westdc.westgis.ac.cn/) (Figure 2). a 0 , b 0 , and c 0 are given parameters, and b 0 , c 0 ∈ (0, 1).Referencing from the study of Lievens et al. [30], a 0 , b 0 , and c 0 are given as 0.02, 0.5, and 0.02, respectively, in this study.It should be noted that when the ESA CCI SM is high orderly rescaled (Section 3.3.3),the observation error parameter σ R needs to be rescaled according to where σ * R is the standard deviation (SD) of the rescaled ESA CCI SM observation error, and σ sim and σ obs are the SD of the simulated SM error and the ESA CCI SM error, respectively.

Evaluation Metrics.
e relative error (RE), the root mean square error (RMSE), the Nash-Sutcliffe coefficient of efficiency (NSE), and Pearson's correlation coefficient (R) are used to measure the coincidence level of the simulated streamflow to the field observations.Meanwhile, the effectiveness criterion (EFF) [55] and the normalized error reduction index (NER) are used to directly assess the performance of soil moisture assimilation.
RE describes the deviation rate (%) of the predicted streamflow to its field measurements.It can be expressed by where n is the total time step and Q sim i and Q obs i are the simulated and observed streamflow at time i.
NSE is expressed by where Q obs indicates the mean value of the measured streamflow for the whole period.is NSE expression puts more importance on high flow.In order to give more weight to low flow, a modified version of the Nash-Sutcliffe coefficient of efficiency is adopted.It is actually a calculation of the NSE in a logarithmic form of the variable (NSE log ): EFF reflects the data assimilation effects by comparing the sum of square error between the streamflow under assimilated and nonassimilated cases.It can be expressed as where Q EnKF i and Q EnOL i are the predicted streamflow under assimilated and nonassimilated cases at time i.
NER is expressed by the following [55]: where

Results
e e ciency of ESA CCI SM assimilation is highly dependent on the quality of model calibration, model, and observation error estimation.Hence, the ESA CCI SM assimilation e ects on stream ow simulation accompanied by the results for model calibration and validation and for model and observation error estimation are analyzed.To illustrate the e ciency of DA, the ensemble open-loop (EnOL) cases and the EnKF cases are compared.e EnOL is an ensemble running of the SWAT model with perturbations on model inputs, model parameters, and model states without the integration of observed SM, while the EnKF is an ensemble running of the SWAT model with the same perturbation to EnOL, but with the integration of ESA CCI SM during the model propagation process. 2 presents the SWAT model parameters being calibrated, which are obtained from the parameter sensitivity analysis detailed in Section 3.3.1.Figure 4 plots the simulated and observed daily series of runo at the catchment outlet during the calibration (1992-1999) and validation (2002)(2003)(2004)(2005)(2006)(2007)(2008) periods.e hydrograph of the simulated stream ow is highly consistent with that of the observed stream ow for both the calibration and validation stages, although slight underestimation exists in ood peak modeling over some periods.

Model Calibration and Validation. Table
e statistics (Table 3) for the simulated stream ow at the catchment outlet (Huaibin) suggest that it agrees well with the measured runo as RE < 5%, NSE > 0.8, and R > 0.9.In addition, the statistics for the other ve hydrological sites (Dapoling, Changtaiguan, Zhuganfu, Xixian, and Huangchuan) also indicate that the SWAT model has fairly good applicability in the upper Huai River basin.In the calibration stage, for all six stations, RE < 15%, NSE falls between 0.65∼0.81,and R > 0.83.In the validation state, RE < 15%, NSE falls between

Model Error Estimation for Precipitation and Model
Parameter.Figure 5 shows the objective function F (11) with varying standard deviation (SD) of the lognormal multiplicative perturbation on precipitation (σ p ) from 0.05 to 0.5 along with the varying SD of the normal multiplicative perturbation on parameter from 0.1 to 0.5.e objective function F reaches its minimum value when σ p approximates 0.35, which suggests that, in this case, the simulated outlet stream ow is the best matching to the observed stream ow from its ensemble statistics.However, it can be seen that σ par is not that sensitive to the objective function, and the allotropism or nonuniqueness issue exits in its optimal parameter estimation.Considering the good performance of the SWATmodel in the study basin (Section 4.1), the small values of σ par (<0.25) are more credible.Besides, in consideration of the robustness of the EnKF method [45], 0.25 for σ par is adopted.Note that, in this estimation of σ p and σ par , the observed stream ow at the subbasin outlet is regarded as the truth, that is, the observation error is ignored.In this case, σ p and σ par are likely to be slightly overestimated.

ESA CCI SM Error Estimation.
Figure 6 shows the standard deviation (SD) of the observation error σ * R (13) for the ESA CCI SM at each grid (34 grids in total) within the catchment.e location for 1-34 grids is present in Figure 7.In general, σ * R falls between 0.03 and 0.05 m 3 /m 3 for all grids, which is consistent with the accuracy of the ESA CCI SM on average (0.04∼0.05 m 3 /m 3 ) [57].At each grid, σ * R presents certain ranges (∼0.01 m 3 /m 3 ), which is related to the soil moisture dynamic with time changes over [2003][2004][2005][2006]. is also indicates the necessity for considering the temporal characteristics of the observation error for RS SM.In addition, σ * R shows a considerable di erence among di erent grids.e high σ * R mainly appears on grids with the dense forest coverage, for example, the grids 10, 11, 25, 26, 33, and 34, in particular for the grid 33. e dense forest obscures the emitted radiance of the soil surface, which results in large uncertainty to the surface SM retrieval.Besides, major dense forests are distributed over the catchment with high altitudes (Figure 1), where the complex topography also impedes the accuracy and reliability of remote sensing for SM [58].

ESA CCI SM Assimilation on Stream ow Simulation.
Table 4 statistically compares the model simulated stream ow with (EnKF) and without (EnOL) ESA CCI SM assimilation at the six hydrologic sites in the upper Huai River basin except for Huangchuan (Figure 1).e reason for Huangchuan not being taken into account is that it has data quality issue over 2004 and 2006 caused by severe human activities.Table 4 shows that the RE and RMSE are decreased and the NSE, NSE log , and R are increased at the ve gauges except for Zhuganfu due to ESA CCI soil moisture assimilation.e improvement is more signi cant in terms of the NSE log as its increase rate is greater than NSE and R, which indicates that the RS soil moisture assimilation is more e ective for low ows than high ows.Besides, EFF/NER > 0 for four sites, in particular for Dapoling, Xixian, and Huaibin (where NER > 5% and EFF > 10%), which suggests the good performance of the assimilation.e none ective performance of ESA CCI SM assimilation on runo simulation of Changtaiguan and Zhuganfu is probably related to their large proportions of the dense forest and complex topography coverage upstream (Figures 1 and 7).Both dense forest coverage and complex topographical conditions reduce the data quality of RS SM retrievals, thus impeding its performance in DA.   and the study of Massari et al. [32].e uncertain performance of soil moisture assimilation on large-ood simulation mainly lies in the relatively low dependence of runo generation on antecedent soil moisture because during large-ood periods, the soil moisture is nearly saturated and the runo is largely controlled by precipitation inputs.

Discussion
In general, our results indicate that the ESA CCI soil moisture assimilation in SWAT performs well in runo modeling of the whole basin.Stream ow improvements over ve in situ sites (except for Zhuganfu) are shown after proper con gurations of the model and observation error.However, the improvements are not signi cant, which can be attributed to the following factors: First, the model error is estimated based on analyzing the ensemble characteristics of the stream ow simulations driven by the model error perturbations in reference to the ground-based runo observations, during which the observation error for the stream ow is ignored.It might lead to an overestimated model error.In observation error estimation for satellite soil moisture, subjectivity does exist in parameter assignment of the estimation equation although the temporal and spatial variability has been taken into account.ese two factors are likely to deteriorate the model and observation error estimation, which eventually degrade the DA performance.Second, the runo improvements are obtained by updating the pro le SM using the satellite SM products, which highly relies on the physical vertical coupling-based model.SWAT soil layers have limited vertical coupling [12,18] as it does not allow actual soil water compensation from other soil layers in the storage routing technique, and the soil water de ciency is only made up by adjusting the depth distribution of soil evaporation demand.e exponential lter [59] used to derive the pro le SM indicator from the surface SM observations is a common solution to the inconsistency of the shallow (surface) RS detection and the runo root zone control mechanism [5,32].However, this approach is more applicable to the hydrological model with a single soil layer setup.For the multilayer setup model (e.g., SWAT), a more promising approach to the physical coupling issue would be adopting Richard's equation because it is more representative of the real-world water movement of soil water.Finally, the runo improvements show large discrepancies over di erent hydrological sites with di erent geographical locations, which suggests that the land surface conditions considerably in uence the DA performance (similar to the results from the study of Massari et al. [32]).e dense forests and complex topographical conditions reduce the data quality of microwave soil moisture retrievals, thus deteriorating the e ciency of satellite soil moisture assimilation.Remote sensing grid Forest coverage

Figure 1 :
Figure 1: e upper Huai River basin: location (the inset map), elevation, digital river network, location of the meteorological and hydrological stations (ST), and the subbasin delineation in SWAT model building.

Figure 2 :
Figure 2: Land use/land cover (LU/LC) classes and soil distribution for the upper Huai River basin.

Figure 8
compares the daily series of the model simulated stream ow at the catchment outlet during 2003-2006 with (EnKF) and without (EnOL) ESA CCI SM assimilation on the basis of the observed runo (Obs).It can be seen that ESA CCI SM assimilation improves the stream ow modeling over low-ow periods.e predicted runo with ESA CCI SM

Figure 5 :
Figure 5: Contour plot of the objective function F (11) with respect to the standard error deviation for precipitation (σ p ) and for model parameter (σ par ).

Figure 6 :
Figure 6: Standard deviation (σ * R ) of the observation error for the rescaled ESA CCI SM at each grid within the basin.e upper limb, the lower limb, and the red line in the box plot represent the upper quartile (h 1 ), the lower quartile (h 2 ), and the median value of σ * R over 2003-2006.e dotted line stretched from the box is the range between h 1 − 1.5 (h 2 − h 1 ) and h 2 + 1.5 (h 2 − h 1 ).

Figure 7 :
Figure 7: Forest coverage for the remote sensing grids of the ESA CCI SM over the upper Huai River basin.1-34 is the grid coding for the ESA CCI SM.

Figure 8 :
Figure 8: e model-simulated stream ow at the catchment outlet (Huaibin) with (EnKF) and without (EnOL) the ESA CCI SM data assimilation over 2003-2006.e upper plot is for the whole period with a base 10 logarithm coordinate.e four plots below are for the ood season from 2003 to 2006.e red and blue lines are the EnKF and EnOL ensemble members, and the red and blue bold lines represent the mean value of EnKF and EnOL.Obs is the abbreviation of observation.

Table 1 :
Soil classification and its area proportions in the upper Huai River basin.P s,t is the error covariance of the measurement prediction at time t, and R s,t is the error covariance of the RS SM at time t.
3.3.Implementation of the ESA CCI SM Assimilation in SWAT.To set up the ESA CCI SM data assimilation in the SWAT model, the SWAT model for the study catchment should be rstly built up.e model applicability is evaluated over the model calibration and validation processes.en, the model error for the validated SWATmodel is estimated based on the in situ stream ow observations at the Huaibin hydrologic station.Besides, the observation error estimation for the ESA CCI SM is implemented before the bias correction on the basis of the model-simulated soil moisture.A ow chart for implementing ESA CCI SM assimilation in SWAT is shown in Figure

Table 2 :
SWAT model parameters being calibrated.

Table 4 :
Statistical comparison of the estimated stream ow in EnOL and EnKF cases based on the observed stream ow over 2003-2006.RMSE (m 3 /s) NSE NSE log R RE (%) RMSE (m 3 /s) NSE NSE log R