The model domain, the Rur catchment (2354 km2), is situated in western Germany and illustrated in Fig. 1. The altitude varies between 15 m a.s.l. in the flat northern part and 690 m a.s.l. in the hilly southern part. Precipitation, evapotranspiration and land use follow the topography. The dominant land use types are agriculture (mainly in the north), grassland, and coniferous and deciduous forest. Annual precipitation ranges between less than 600 mm in the north and 1200 mm in the hilly south (Montzka et al., 2008). Annual potential evapotranspiration varies between 500 mm in the south and 700 mm in the north (Bogena et al., 2005). The Rur catchment CRNS network comprises nine CRNSs (CRS1000, HydroInnova LLC, 2009) which were installed in 2011 and 2012 (Baatz et al., 2014). Climate and soil texture of the CRNS sites can be found in Table 1.

The CRNSs were calibrated in the field using gravimetric soil samples. At each site, 18 soil samples were taken along 3 circles with distances of 25, 75 and 175 m from the CRNS, and 6 samples were evenly distributed along each circle. Each sample was extracted with a 50.8 × 300 mm round HUMAX soil corer (Martin Burch AG, Switzerland). The samples were split into 6 sub-samples with 5 cm length each and oven dried at 105 ∘C for 48 h to measure dry soil bulk density and soil moisture. Lattice water, hydrogen from organic and an-organic sources, was determined for each site using a heat conductivity detector (Ray, 1954). Soil bulk density, soil moisture, lattice water and 12 h averaged measured neutron intensity were used to determine calibration parameters specific to each CRNS and the COSMIC operator. This represents a compromise between the measurement noise (which follows a Poisson distribution) and the assumed variation of environmental variables over the averaging time window (Iwema et al., 2015).

The CLM was the land surface model of choice for simulating water and energy exchange between the land surface and the atmosphere (Oleson et al., 2013). Some of the key processes which are modelled by CLM are radiative transfer in the canopy space, interception of precipitation by the vegetation and evaporation from intercepted water, water uptake by vegetation and transpiration, soil evaporation, and photosynthesis, as well as water and energy flow in the subsurface. SWC in CLM is influenced by precipitation, infiltration into the soil, water uptake by vegetation, surface evaporation, and surface and subsurface runoff. To limit the scope and complexity of this study, CLM was run using satellite phenology, e.g. prescribed leaf area index data and the biogeochemical module turned off. The biogeochemical module allows CLM to model the vegetation development dynamically, but it requires a large spin-up of 1000 years, and little additional gain is expected for this study from these additionally modelled processes.

Vertical water flow in soils is modelled by the 1-D Richards equation. Soil hydraulic parameters are determined from sand and clay content using pedotransfer functions for the mineral soil fraction (Clapp and Hornberger, 1978; Cosby et al., 1984) and organic matter content for the organic soil fraction (Lawrence and Slater, 2008).

The joint state–parameter estimation used in this study updates soil texture and organic matter in CLM. Hence, parameter estimates directly determine soil hydraulic properties in CLM. The following equations describe how soil texture and organic matter define the soil hydraulic properties in CLM such as porosity, hydraulic conductivity, the empirical exponent B and soil matric potential. Hydraulic conductivity (k(zi), mm s-1) at the depth z between two layers (i and i + 1) is a function of soil moisture (θ, m3 m-3 in layers i and i + 1), saturated hydraulic conductivity (ksat(zi), mm s-1), saturated soil moisture (θsat, m3 m-3) and the empirical exponent B (Oleson et al., 2013): kzi=ϕiceksatziθi+θi+1θsat,i+θsat,i+12Bi+3,1≤i≤Nlevsoi-1ϕiceksatziθiθsat,i2Bi+3,i=Nlevsoi, where ϕice is the ice impedance factor. The ice impedance factor was implemented to simplify an increased tortuosity of water flow in a partly frozen pore space. It is calculated with ϕice = 10-ΩFice using the resistance factor Ω = 6 and the frozen fraction of soil porosity Fice = θice/θsat,i. Soil hydraulic properties are calculated separately for the mineral (min) and organic matter (om) soil components. Total porosity θsat,i is calculated using the fraction of organic matter (fom,i) with the following: θsat,i=1-fom,iθsat,min,i+fom,iθsat,om, where the organic matter porosity is θsat,om = 0.9 and sand content in percentage (%) determines the mineral soil porosity θsat,min as follows: θsat,min=0.489-0.00126×%sand. Analogous, the exponent B is calculated with Bi=1-fom,iBmin,i+fom,iBom, where Bom = 2.7 is the organic exponent and the mineral exponent Bmin,i is determined by clay content in percentage (%) with the following: Bmin,i=2.91+0.159×%clay. Saturated hydraulic conductivity is calculated for a connected and an unconnected fraction of the grid cell with the following: ksatzi=1-fpercksat,unconzi+fperc,iksat,omzi, where fperc,i is the fraction of a grid cell where water flows with saturated hydraulic conductivity of the organic matter (ksat,om(zi) in mm s-1) through the organic material only, the so-called connected flow pathway. The saturated hydraulic conductivity of the unconnected part (ksat,uncon(zi), mm s-1) depends on organic and saturated mineral soil hydraulic conductivity: ksat,uncon=1-fperc1-fomksat,min+fom-fpercksat,om-1, where saturated hydraulic conductivity for mineral soil is calculated from the grid cell sand content as follows: ksat,minzi=0.0070556×10-0.884+0.0153×%sand. The fraction fperc is calculated with the following: fperc=0.908×fom-0.50.139,fom≥0.50,fom<0.5. Soil matric potential (mm) is defined as a function of saturated soil matric potential (mm) with the following: ψi=ψsat,iθiθsat,i-Bi=1-fom,iψsat,min,i+fom,iψsat,omθiθsat,i-Bi, where saturated organic matter matric potential is ψsat,om = -10.3 mm and saturated mineral soil matric potential is calculated from sand content as follows: ψsat,min,i=-10.0×101.88-0.0131×%sand

SWC retrievals were calculated from neutron intensity observations with COSMIC (Shuttleworth et al., 2013) following calibration results and the procedure of Baatz et al. (2014). COSMIC parameterizes neutron transport within the soil subsurface and was calibrated against the more complex Monte Carlo Neutron Particle model MCNPx (Pelowitz, 2005). COSMIC needs considerably less CPU time than the MCNPx model. The code was tested at multiple sites for soil moisture determination (Baatz et al., 2014; Rosolem et al., 2014) and analysed in detail by Rosolem et al. (2014).

COSMIC assumes that a number of high-energy neutrons enter the soil. In the soil, the number of high-energy neutrons is reduced by interactions within the soil, leading to generation of fast neutrons in each soil layer. Before resurfacing, the number of fast neutrons is reduced again by their interaction with nuclei of elements within soil (Shuttleworth et al., 2013). The number of neutrons, NCRP, that reaches the CRNS can be summarized in a single integral as follows: NCRP=NCOSMIC∫0∞A(z)αρS(z)+ρw(z)exp⁡-ms(z)L1+mw(z)L2⋅dz, where NCOSMIC is an empirical coefficient that is CRNS-specific and needs to be estimated by calibration, A(z) is the integrated average attenuation of fast neutrons, α = 0.404 - 0.101 × ρS is the site-specific empirical coefficient for the creation of fast neutrons by soil, ρS is the dry soil bulk density (g cm-3), ρw is the total soil water density (g cm-3), and ms and mw are the masses of soil and water, respectively, per area (g cm-2) L1 = 162.0 g cm-2 and L2 = 129.1 g cm-2 are scattering lengths for fast neutrons in solids and water, respectively, that were estimated using the MCNPx code (Shuttleworth et al., 2013). The integrated average attenuation of fast neutrons A(z) can be found numerically by solving A(z)=2π∫0π/2exp⁡-1cos⁡(γ)ms(z)L3+mw(z)L4⋅dγ, where γ is the angle along a vertical line below the CRNS detector to the element that contributes to the attenuation of fast neutrons, and L3 = -31.65 + 99.29 × ρS and L4 = 3.16 g cm-2 are the scattering lengths for fast neutrons in soil and water, respectively, determined using the MCNPx code (Shuttleworth et al., 2013). The COSMIC operator is discretized into 300 layers of 1 cm thickness up to a depth of 3 m. For each CLM grid cell in the model domain, simulated SWC in all CLM layers is used to generate a weighted SWC retrieval using the COSMIC code. Simulated SWC is handed from the CLM simulation history files to the COSMIC operator. Given the vertical SWC distribution of the individual CLM grid cell, COSMIC internally calculates the contribution of each layer to the simulated neutron intensity signal at the soil surface in COSMIC. In this study, the contribution of each CLM soil layer was used to calculate the weighted CLM SWC retrieval corresponding to the vertical distribution of simulated SWC in each grid cell.

Measured neutron intensity of CRNS was used to inversely determine a CRNS SWC retrieval, as by Baatz et al. (2014) assuming a homogeneous vertical SWC distribution. Then, the weighted CLM SWC retrieval is used in the data assimilation scheme to relate the CRNS SWC retrieval to the model state. Alternatively, neutron flux data could be assimilated directly within the catchment. This would require calibration data throughout the catchment, which is only feasible using spatially distributed data sets (e.g. Avery et al., 2016). However, high stands of biomass are a major factor for calibration in the Rur catchment (Baatz et al., 2015), and estimates of biomass come along with high uncertainties. To circumvent the introduction of these additional uncertainties, SWC retrievals are assimilated in this study. Changes in on-site biomass were assumed to be negligible.

To further expand the work of Han et al. (2016), this study uses the LETKF (Hunt et al., 2007) to assimilate SWC retrievals by CRNSs into the land surface model CLM. Updates were calculated either for SWC states or jointly for SWC states and soil parameters, depending on the experiment setup. For state updates only, the LETKF was used as proposed by Hunt et al. (2007). Calculations were made for an ensemble of model simulations which differed depending on variations in model forcings and input parameters. The states of the different ensemble members are indicated by xif where i = 1, …, N and N is the number of ensemble members; “f” marks the model prediction or forecast before the update. The individual state vectors xif contain the CLM-simulated SWC of the 10 soil layers and the vertically weighted SWC retrieval obtained with the COSMIC operator. For each grid cell, a matrix Xf can be constructed which contains the deviations of the simulated states with respect to the ensemble mean x‾f: Xf=x1f-x‾f,…,xNf-x‾f.

In the case of joint state–parameter updates, a state augmentation approach was followed (Hendricks Franssen and Kinzelbach, 2008; Han et al., 2014). In this case, the augmented model state matrix Xf is constructed from the simulated SWC of the 10 soil layers, weighted SWC, and the grid cell's sand, clay and organic matter content.

In order to relate the measured neutron intensity with the simulated SWC of CLM, the observation operator H (COSMIC) is applied on the measured neutron intensity in order to obtain the expected weighted SWC retrieval at each of the observation locations for each of the stochastic realizations: yif=Hxif. The ensemble realizations of the modelled SWC retrievals at the measurement locations y1f to yNf with respect to the ensemble mean y‾f are stored in the matrix Yf: Yf=y1f-y‾f,…,yNf-y‾f. The observation error correlation was reduced in space by the factor fred using the spherical model: fred=1-1.5×d/dmax+0.5×d/dmax3, where d is the distance to the observation and dmax = 40 km is the maximum observation correlation length, about half the size of the catchment. Only SWC retrievals within the maximum observation correlation length were used for assimilation. This leads to a “localized” size of Yf and the observation error covariance matrix R. The intermediate covariance matrix Pa (also called analysis error covariance matrix) is calculated according to the following: Pa=(N-1)I+YfTR-1Yf. In addition, the mean weight vector w‾a is obtained as follows: w‾a=PaYfTR-1y0-y‾f, where y0 contains the CRNS SWC retrievals at the measurement locations. In the ensemble space, a perturbation matrix Wa is calculated from the symmetric square root of Pa: Wa=(N-1)Pa1/2. The final analysis Xa is obtained from the following: Xa=x‾f+Xfw‾a+Wa. A more detailed description of the LETKF can be found in Hunt et al., (2007), and details on the implementation of the LETKF in combination with CLM are given by Han et al. (2015).

In this study, discretization and parameterization of the hydrological catchment was done on the basis of high-resolution data. The model of the Rur catchment was spatially discretized by rectangular grid cells of 0.008∘ size (∼ 750 m). The model time step was set to hourly. Land cover was assumed to consist of vegetated land units only, and a single plant functional type (PFT) for each grid cell was defined. The plant functional types were derived from a remotely sensed land use map using RapidEye and ASTER data with 15 m resolution (Waldhoff, 2012). Contents of sand, clay and organic matter were derived from the high-resolution regional soil map BK50 (Geologischer Dienst Nordrhein-Westfalen, 2009). Alternative simulations were also performed with the FAO soil map of the global Harmonized World Soil Database (FAO, 2012) and with a biased soil texture with a fixed sand content of 80 % and clay content of 10 % (S80 soil map). Average sand and clay content are 22.5 and 21.4 % for the BK50 soil map and 39 and 22 % for the FAO soil map, respectively. The FAO soil map and the biased soil map represent large error with respect to the soil properties of the BK50 soil map. The FAO soil map and S80 soil map simulations allow the evaluation of the joint state–parameter estimation approach because, given the expected bias, we can evaluate to what extent the soil properties are modified by the data assimilation. This is important because in many regions across the Earth a high-resolution soil map is not available. Land surface models are applied for those regions, for example in the context of global simulations, and hence might be strongly affected by the error in soil properties.

Maximum saturated fraction, a surface parameter which is used for runoff generation, was calculated from a 10 m digital elevation model (scilands GmbH, 2010). Leaf area index data were derived from monthly-averaged Moderate Resolution Imaging Spectrometer data (MODIS). CLM was forced with hourly atmospheric data from the COSMO_DE reanalysis data set for the years 2010 to 2013 from the German Weather Service (Deutscher Wetterdienst, DWD). The data were downscaled from a resolution of 2.8 km × 2.8 km to the CLM resolution using linear interpolation based on Delaunay triangulation. Forcing data include precipitation, incident solar and longwave radiation, air temperature, air pressure, wind speed, and relative humidity at the lowest atmospheric level.

Uncertainty was introduced into the regional CLM model by perturbed soil parameters and external model forcings. Contents of sand, clay and organic matter were perturbed with spatially correlated noise from a uniform sampling distribution with mean zero and standard deviations 10 and 30 % (Han et al., 2015). Soil texture perturbation considers that in CLM a single set of pedotransfer functions is assumed to be valid throughout the globe while pedotransfer functions are usually specific to regions (e.g. Patil and Singh, 2016). In other words, the perturbation of soil texture also covers the uncertainty in the pedotransfer function itself. By perturbing texture, soil parameters are also perturbed through the pedotransfer functions used in CLM as specified in Sect. 2.2. Precipitation (σ = 0.5 or 1.0; lognormal distribution) and shortwave radiation (σ = 0.3; lognormal distribution) were perturbed with multiplicative noise with mean equal to 1. Longwave radiation (σ = 20 W m-2) and air temperature (σ = 1 K) were perturbed with additive noise. The forcing perturbations were imposed with correlations in space (5 km) using a fast Fourier transform. Correlation in time was introduced with an AR(1) model with autoregressive parameter 0.33. These correlations and standard deviations were chosen based on previous data assimilation experiments (Reichle et al., 2010; Kumar et al., 2012; De Lannoy et al., 2012; Han et al., 2015). In this work, only results for precipitation perturbation with σ = 0.5 will be shown, as results for σ = 1.0 were similar. An ensemble size of 95 realizations was used in the simulations. Based on previous work (Baatz et al., 2015), the SWC retrieval uncertainty for CRNS was estimated to be 0.03 cm3 cm-3 while fluctuations in the measurement standard deviation, related to the nonlinear relation between observed neutron intensity and SWC, were assumed to be negligible.

All simulation experiments in this study used initial conditions from a single 5-year spin-up run in which a single forcing data set of the year 2010 was repeatedly used as atmospheric input. The soil moisture regime became stable after the 5-year spin-up period, and additional spin-up simulations would not affect soil moisture in the consecutive years. After this 5-year spin-up, soil parameters and forcing data of the consecutive years were perturbed. From 1 January 2011 onwards, CLM was propagated forward with an ensemble of 95 realizations. On 20 March 2011, the first SWC retrieval was assimilated, and assimilation of SWC retrievals continued until 31 December 2012. In the data assimilation period, soil properties were estimated at every time step when observations were made available. For the year 2013, the model was propagated forward without data assimilation but with an ensemble of 95 realizations. The year 2013 was used exclusively as the evaluation period for data assimilation experiments.

In total, 31 simulation experiments were carried out using different setups (Table 2). The present setups are intended to cover three different initial soil maps, three different sizes of a CRNS network and two different parameter perturbations. Three open loop simulations were run without data assimilation and soil parameter perturbation of 30 % for the BK50 soil map (OL-BK50), the FAO soil map (OL-FAO) and the S80 soil map (OL-S80). These simulations are referred to as reference runs for the respective soil map. Simulation results of data assimilation runs were compared to the reference runs for quantification of data assimilation benefits. Simulations were done with joint state–parameter estimation (PAR-*), two for the BK50 soil map (PAR-BK50-*), one for the FAO soil map (PAR-FAO-30), and two for the S80 soil map (PAR-S80-*). Soil texture was perturbed by 10 or 30 % as indicated by the experiment name (Table 2). Two simulations were done with state updates only for the BK50 soil map (Stt-BK50) and the S80 soil map (Stt-BK50). These 10 simulations form the basic set of experiments.

Besides the data assimilation experiments, a larger number of jackknifing simulations were also conducted to evaluate the impact of the CRNS data assimilation on SWC at unobserved locations in the model domain. In nine jackknife experiments, data from eight CRNS locations were assimilated (jk8-* simulations) and data of the one remaining CRNSs were not assimilated but kept for evaluation. In addition, three simulations were conducted where data of four CRNSs were assimilated (jk4-* simulations), and data of the five remaining CRNSs were used for evaluation. These three simulations represent a CRNS network with much less than the existing nine CRNSs. At the evaluation locations, simulated SWC (which is affected by the assimilation of the other eight probes) was compared to CRNS SWC retrievals. For jackknife simulations, the perturbation of soil texture was set to 30 %. States and parameters at these sites were jointly updated, and simulations were made using either the BK50 or the S80 soil maps as initial parameterization. Therefore, a total of 21 jackknife simulations were performed.

Simulation results were evaluated with the root mean square error (ERMS): ERMS=∑t=1nθt,CLM-θt,CRNS2n, where n is the total number of time steps, θt,CLM is SWC simulated by CLM at time step t and θt,CRNS is the CRNS SWC retrieval at time step t. In case SWC was assimilated at the corresponding time step, θt,CLM is SWC prior to assimilation. In case the ERMS is estimated at a single point in time over all CRNSs available, the number of time steps n can be replaced by the number of CRNSs available. The second evaluation measurement in this study is the bias which is, in contrast to the ERMS, a measure for systematic deviation: bias=∑t=1nθt,CLM-θt,CRNSn.

Table 3 summarizes the performance statistics in terms of ERMS and bias for the assimilation period (2011 and 2012) and evaluation period (2013). Presented are results for the open loop scenarios with the BK50, FAO and S80, and data assimilation scenarios. Errors of open loop simulations were highest for the S80 simulation (0.11 cm3 cm-3), followed by the FAO simulation (0.07 cm3 cm-3) and the BK50 simulation (0.04 cm3 cm-3). Mean absolute bias was highest for the S80 soil map (0.11 cm3 cm-3), now as high for the FAO soil map (0.06 cm3 cm-3) and lowest for the BK50 soil map (0.02 cm3 cm-3). Data assimilation improved simulations more for the S80 soil map (ERMS reduced by 0.08 cm3 cm-3) than for the FAO soil map (ERMS reduced by 0.04 cm3 cm-3) or the BK50 soil map (ERMS reduced by 0.01 cm3 cm-3). The BK50 soil map led to ERMS values in open loop simulations lower than 0.05 cm3 cm-3, which left little room for error reduction considering a measurement error of 0.03 cm3 cm-3. However, slight improvements by 0.01 cm3 cm-3 were possible at monitored locations in the data assimilation period but not in the evaluation period. Joint state–parameter estimation improved simulation results, as shown by the reduced ERMS and bias for the S80 and the FAO soil maps. The verification period (2013) with the updated soil hydraulic parameters for the FAO soil map resulted in an ERMS value of 0.05 cm3 cm-3, also clearly an improvement compared to the open loop run with an ERMS of 0.07 cm3 cm-3 (Table 3). Joint state–parameter updating resulted in similar ERMS values for all three initial soil maps: the BK50, FAO and S80 soil maps (each 0.03 cm3 cm3). State updates (Stt-S80) improved ERMS and bias for the S80 soil map (ERMS = 0.06 cm3 cm-3 for assimilation period) but much less compared to the joint state–parameter updates (PAR-S80-30). The ERMS and bias for simulations with 10 and 30 % perturbation of soil texture only showed very small differences (smaller than 0.01 cm3 cm-3).

The temporal course of simulated soil moisture in 2011 at the two sites, Merzenhausen and Gevenich, is shown in Fig. 2. The figure illustrates that simulated SWC at both sites was lowest with the S80 soil map (OL-S80) and highest with the BK50 soil map (OL-BK50), and the FAO soil map resulted in intermediate soil moisture (OL-FAO). Mean open loop SWC in 2011 was 0.17 cm3 cm-3 for the S80 soil map, 0.24 cm3 cm-3 for the FAO soil map and 0.27 cm3 cm-3 for the BK50 soil map at both sites. Measurements with CRNS started in May 2011 at Merzenhausen. At Gevenich, the first observation was recorded on 7 July 2011. In the data assimilation run shown (PAR-S80-30), modelled SWC was immediately affected at both sites, Merzenhausen and Gevenich, as soon as data at Merzenhausen were assimilated. By July, simulated SWC with the biased soil map and data assimilation (PAR-S80-30) was already close to the CRNS SWC retrieval at the Gevenich site (Fig. 2). This demonstrates the beneficial impact of data availability for assimilation at one site and the information brought into space by the data assimilation scheme. Figure 2 also shows that the BK50 open loop run was close to the observed SWC at both sites, even without data assimilation.

Figure 3 shows the temporal course of SWC from January 2011 to December 2013 at Heinsberg and Wildenrath. Assimilation and evaluation results are shown for the open loop (OL-S80 and OL-FAO) simulations, only state updates (Stt-S80), joint state–parameter updates (PAR-S80-30) and CRNS SWC retrievals. At Heinsberg, results show that simulated SWC with assimilation was closer to the CRNS when both states and parameters were updated (PAR-S80-30) than if only states were updated (Stt-S80). This is the case for both the assimilation and the evaluation periods. At the beginning of the evaluation period (first few days of 2013), the Stt-S80 simulation shows an increase in bias between modelled SWC and CRNS. The bias of Stt-S80 remained throughout the evaluation period. In contrast, if parameters were previously updated (PAR-S80-30), modelled SWC was close to the CRNS during the evaluation period. Open loop SWC modelled with the FAO soil map is lower than the CRNS SWC retrievals at Heinsberg and higher than CRNS SWC retrievals at Wildenrath. At Wildenrath, results of the OL-S80 run suggest that the initial sand content of the biased soil map is closer to the optimal sand content than the sand content of the FAO soil map. Consequently, the OL-FAO bias was -0.05 and 0.05 cm3 cm-3 for Heinsberg and Wildenrath, respectively (Tables A1 and A4 in Appendix). At both sites, absolute bias was reduced with joint state–parameter updates to equal or less than 0.01 cm3 cm-3 (S80 and FAO soil map). The reduced bias is also well reflected in the temporal course of modelled SWC with joint state–parameter updates (PAR-S80-30).

It is interesting to notice that the error values for the verification period are very similar if soil hydraulic parameters were estimated in the assimilation period, independent of the initial soil map (Table 3). ERMS values for the 2013 simulations with state updates only (Stt-BK50 and Stt-BK50) show that in the evaluation period the improvements by state updates (without parameter updates) were small (reduction by 0.02 and 0.00 cm3 cm-3 for S80 and BK50, respectively) compared to the improvements obtained by joint state–parameter updates (reduction by 0.07 cm3 cm-3 for S80). This illustrates the benefits of joint state–parameter updates compared to state updates only, and that soil moisture states are strongly determined by soil hydraulic parameters. The case of only state updates also illustrates that the improved characterization of soil moisture states in the assimilation period results in improved initial states for the verification period (Table 3) but in the verification period these improvements lose their influence quickly over time (Fig. 3).

Figure 4 shows the temporal evolution of the hourly ERMS calculated for all nine CRNSs. ERMS was highest for the S80 open loop run and lowest for the PAR-S80-30 simulation. The FAO soil map resulted in errors mostly between 0.05 and 0.1 cm3 cm-3, which are lower than the S80 soil map but not as good as simulation results with joint state–parameter updates (PAR-S80-30) or with the BK50 soil map (OL-BK50). State updates did not improve modelled SWC as much as joint state–parameter updates. For most of the time, the ERMS of the Stt-S80 run is larger than the ERMS of the OL-BK50 run. During the evaluation period, the open loop run with the FAO soil map (OL-FAO) also performs better than the Stt-S80 run. In contrast, joint state–parameter updates to the S80 soil map improved the ERMS most of the time compared to open loop simulations (OL-BK50, OL-FAO and OL-S80). As shown in Fig. 4, the PAR-S80-30 simulation performed best out of the four simulations during the assimilation period 2011–2012. During the evaluation period 2013, OL-BK50 and PAR-S80-30 performed equally well, except in summer 2013 when the PAR-S80-30 simulation yielded much higher ERMS values than the BK50 open loop run.

The jackknife simulations investigated the impact of CRNS data on improving simulated SWC at locations beyond the CRNS stations. Spatial improvements are possible by spatial correlation structures of atmospheric forcings, soil hydraulic parameters and soil moisture which are taken into account by the local ensemble transform Kalman filter. The error and bias shown in Table 4 refer to jackknife simulations with the BK50 and the S80 soil map. On average, over the three runs where only data of four CRNSs were assimilated (jk4-S80-*), the ERMS was 0.07 m3 m-3, which is much lower than the ERMS for the open loop run (0.12 m3 m-3) and only a bit higher than the case where eight CRNSs were assimilated (ERMS = 0.06 m3 m-3 for jk8-S80-*). The improved simulation results were also due to the bias reduction from 0.11 to 0.05 m3 m-3 in the case of four and 0.04 m3 m-3 in the case of eight assimilated CRNSs. However, for the BK50 soil map where ERMS (0.04 m3 m-3) and bias (0.02 m3 m-3) of the open loop run were already good, the jackknife simulations led to slightly higher ERMS (0.05 m3 m-3) and bias (0.04 m3 m-3). More detailed site statistics (Tables A1 to A4 of the Appendix) demonstrate that all jackknife simulations with the S80 soil map resulted in an improved ERMS at the omitted locations compared to the open loop simulation, except for Wildenrath. At sites with large open loop ERMS, the assimilation could reduce the ERMS by 50 % or more.

The jackknife simulations illustrate that a network of CRNSs can improve modelled SWC if the soil map information is not sufficient. This suggests that assimilation of CRNS data is particularly useful for regions with little information on subsurface parameters. A trade-off can be expected between the initial uncertainty on soil moisture and parameters, and the density of a CRNS network. In the case of a large uncertainty, like in regions with limited information about soils or a strongly biased soil map (e.g. FAO or S80 soil map) and a low density of meteorological stations, a sparse network of probes can already be helpful for improving soil moisture characterization. The results of the real-world jackknife experiments demonstrated that four CRNSs are already beneficial, but it is desirable to have more CRNSs for improved parameter estimates. The results also suggest that the additional information gain for an extra CRNS is reduced for a denser network, because the soil moisture characterization did not improve so much more when eight instead of four CRNSs were used for assimilation. However, in regions with a high density of meteorological stations and a high-resolution soil map, it can be expected that a denser CRNS network than that in this study is needed to further lower the error of soil moisture characterization. Further potentially synthetic experiments in other regions with networks of CRNSs are needed to obtain more quantitative information about this.

The temporal evolution of sand content estimates during the assimilation period for the nine sites with CRNSs is shown in Fig. 5 for PAR-S80-30, PAR-S80-10, PAR-BK50-30, PAR-BK50-10, jk8-S80-* and jk8-BK50-*. Time series start on 20 March 2011, the date of the first assimilated CRNS SWC retrieval at Wüstebach. At Wüstebach and sites within the influence sphere of Wüstebach (Aachen, Kall and Rollesbroich), sand content estimates were updated from 20 March 2011 onwards. Because of the localization, all other sites show a first update in sand content in May 2012 when Rollesbroich and Merzenhausen start operating, and their data were assimilated. During the data assimilation period with joint state–parameter updates, all sites show variability in sand content estimates over time with differences in magnitude. Values and spread in sand content estimates amongst the experiments is smaller at the sites Merzenhausen, Gevenich, RurAue, Heinsberg and Wildenrath, compared to the sites Wüstebach, Aachen and Rollesbroich where spread is considerably larger. At the sites Merzenhausen, Kall, Gevenich, RurAue and Heinsberg, sand content estimates of the jackknife simulations were close to the sand content of the other data assimilation experiments with joint state–parameter estimation. A comparison of parameter estimates at the end of the assimilation period indicates that initial soil parameterization has a limited effect on the resulting parameter estimates. Parameter estimates of jk8-BK50-* and jk8-S80-* are close together at the end of the assimilation period.

Estimates of the soil hydraulic parameter B and saturated hydraulic conductivity are shown in Figs. 6 and 7 for PAR-S80-30, PAR-S80-10, PAR-BK50-30, PAR-BK50-10, jk8-S80-* and jk8-BK50-*. Updates of soil hydraulic parameters start in March and May 2011 with the assimilation of CRNS SWC retrievals, depending on the location. The B parameter estimates increase for all simulations. Throughout the whole assimilation period the empirical B parameter varies considerably within short time intervals. The total range of the B parameter estimates is between 2.7 and 14 at all sites. At the sites Merzenhausen, Kall, Aachen, Gevenich and Rollesbroich, B generally ranges between 6 and 10. At Wüstebach, Heinsberg and RurAue, estimates of B range most of the time between 8 and 12, and at Wildenrath, B is below 8. Initial saturated hydraulic conductivity ksat is rather high (ksat > 0.015 mm s-1) in the case of high sand content, i.e. for the S80 soil map, and rather low (ksat < 0.005 mm s-1) in the case of low sand content, i.e. for the BK50 soil map. In the case of the S80 soil map, at all sites except Wildenrath, high initial ksat estimates decrease quickly through joint state–parameter updates to values below 0.01 mm s-1. The initial spread in ksat estimates amongst the simulation scenarios decreases at most sites. At Wüstebach, Merzenhausen, Aachen, Gevenich, RurAue and Heinsberg, the spread is rather small, particularly at the end of the assimilation period, while at Wildenrath ksat ranges from 0.005 to 0.015 for individual experiments at the end of the assimilation period.

Temporally unstable parameter estimates imply that there may be multiple or seasonal optimal parameter values. This is also supported by the findings of the temporal behaviour of site-average ERMS (Fig. 4), e.g. during the evaluation period when, in the dry summer of 2013, the ERMS peaks for the PAR-S80-30 simulation. In this context, it is important to mention that many possible error sources were not subject to calibration in this study but could be crucial for an even better modelled soil moisture and more reliable soil parameter estimation. In this study we only considered uncertainty of soil parameters, but vegetation parameters are also uncertain. Also, a number of other CLM-specific hydrologic parameters (e.g. decay factor for subsurface runoff and maximum subsurface drainage) strongly influence state variables in CLM and hence show potential for optimization (Sun et al., 2013). Considering this uncertainty from multiple parameters could give a better parameter-uncertainty characterization (Shi et al., 2014). Precipitation is also an important forcing for hydrologic modelling. For this study, precipitation data from the COSMO_DE re-analysis were used. A product which optimally combines precipitation estimates from radar and gauge measurements is expected to give better precipitation estimates than the reanalysis. This could improve the soil moisture characterization and also potentially lead to better parameter estimates. Further improvements and constraining of parameter uncertainty is also possible using multivariate data assimilation with observations such as latent heat flux (e.g. Shi et al., 2014). Also, other error sources related to the model structure play a significant role. These options should be subject to future investigations.

Latent heat flux, or evapotranspiration (ET), is another important diagnostic variable of land surface models (e.g. Best et al., 2015) and is of importance for atmospheric models. Results of the data assimilation experiments showed that soil texture updates altered soil moisture states significantly. In Fig. 8 it is shown that joint state–parameter estimation also altered ET during the evaluation period. Figure 8 shows ET within the evaluation period 2013 across the whole catchment for four simulation experiments. On the one hand, ET was similar for both open loop simulations (OL-S80 and OL-BK50) in the south of the catchment. On the other hand, ET in the north was up to 80 mm lower per year for the S80 open loop run compared to the BK50 open loop run. The differences can be linked to the drier soil conditions for OL-S80 compared to OL-BK50 simulation results. The differences in ET between the runs with and without parameter updates were larger for the S80 soil map than for the BK50 soil map. For PAR-S80-10, ET increased by up to 40 mm yr-1 in the northern part of the catchment through data assimilation while the change in ET from OL-BK50 to PAR-BK50-10 is rather small. This is linked to the comparatively larger updates made to soil hydraulic parameters.

Additionally, the impact of soil parameter estimates on ET is different in the north of the catchment compared to the south. While ET in the north of the catchment was impacted by the estimated soil properties during the evaluation period in 2013 for PAR-S80-10, ET in the south was not impacted as much by estimated soil properties. This is related to the fact that in the north ET is moisture-limited in summer, whereas in the south this is not moisture-limited but energy-limited. Therefore, ET in the north is sensitive to variations in soil hydraulic parameter values, whereas in the south this is not the case. In the south, ET is sensitive to model forcings such as incoming shortwave radiation. Nearing et al. (2016) came to the conclusion that soil parameter uncertainty dominates soil moisture uncertainty and forcing uncertainty dominates ET uncertainty. Our findings in the southern part of the catchment support their conclusion, but in the northern part of the catchment soil parameter uncertainty strongly affects ET. Hence, particularly in the northern part of the catchment, further observations such as ET measurements are desirable for further improving the land surface model. These additional observations could be used for future land surface model benchmarking (Best et al., 2015) or for more constrained parameter estimates (Shi et al., 2015).