A soil moisture and temperature network for SMOS validation in Western Denmark

The Soil Moisture and Ocean Salinity Mission (SMOS) acquires surface soil moisture data globally, and thus product validation for a range of climate and environmental conditions across continents is a crucial step. For this purpose, a soil moisture and temperature network of Decagon ECH2O 5TE capacitance sensors was established 5 in the Skjern River Catchment, Denmark. The objectives of this article are to describe a method to implement a network suited for SMOS validation, and to present sample data collected by the network to verify the approach. The design phase included (1) selection of a single SMOS pixel (44 × 44 km), which is representative of the land surface conditions of the catchment and with minimal impact from open water (2) arrange- 10 ment of three network clusters along the precipitation gradient, and (3) distribution of the stations according to respective fractions of classes representing the prevailing environmental conditions. Overall, measured moisture and temperature patterns could be related to the respective land cover and soil conditions. Texture-dependency of the 0–5 cm soil moisture measurements was demonstrated. Regional di ﬀ erences 15 in 0–5 cm soil moisture, temperature and precipitation between the north-east and south-west were found to be small. A ﬁrst comparison between the 0–5 cm network averages and the SMOS soil moisture (level 2) product is in range with worldwide validation results, showing comparable trends for SMOS retrieved/initial soil moisture and initial temperature ( R 2 of 0.49/0.67 and 0.97, respectively). While retrieved/initial 20 soil moisture indicate signiﬁcant under-/overestimation


Introduction
The assessment of water resources is vital under changing climate and land use, especially when coupled with a steadily increasing population (e.g. FAO-AQUASTAT, 2003). Climate and hydrological models constitute important tools for such investigations, but their reliability is constrained due to uncertainty in important input parameters. One 5 of the key variables is soil moisture, as it significantly impacts water and energy exchanges at the land surface-atmosphere interface, and it represents the main source of water for agriculture and natural vegetation. However, soil moisture is highly variable in space and time and across scales, as a result of spatial heterogeneity in soil and land cover properties, topography and climatic drivers (Famiglietti et al., 1998;Mo-10 hanty et al., 2000;Western et al., 2002) rendering it very difficult to assess. Thus, observations of soil moisture at appropriate spatial and temporal scales are urgently needed.
The Soil Moisture and Ocean Salinity satellite (SMOS, Kerr et al., 2001) is the first mission dedicated to soil moisture measurements. A multi-angle, fully polarimetric (Champagne et al., 2010); Australia (Walker et al., 2001;Merlin et al., 2008); Africa (de Rosnay et al., 2009);Europe -Spain (Martinez-Fernandez and Ceballos, 2003), France , Germany (Krauss et al., 2010;Bogena et al., 2010). Many of them can be found in the International Soil Moisture Network Database (Dorigo et al., 2011). These networks often face constraints with respect to their density or 5 spatial extent (Cosh et al., 2004). Various upscaling techniques have evolved to derive spatial patterns at large scales, e.g. interpolation (Bardossy and Lehmann, 1998), time/rank stability (Vachaud et al., 1985), statistical transformation (Reichle and Koster, 2004;De Lannoy et al., 2007), and land surface modeling (Crow et al., 2005). However, these methods are sometimes themselves vulnerable to coarse spacing or limited 10 extent of in situ data, often requiring costly long-term pre-studies. Methods to a priori design networks in a spatially representative manner would be beneficial. Friesen et al. (2008) presented an approach of area-weighted sampling by means of landscape units (hydrotopes) with internally more consistent hydrologic behavior, whereby variance and bias in the large-scale in situ soil moisture average can be reduced. The method was 15 successfully applied in two short-term campaigns in West Africa. However, it is both region-dependent and quite complex.
Several studies have focused on the number of samples required to estimate the satellite footprint-scale mean. It was noted that soil moisture variability increases with the spatial extent of a footprint, implying an increase in the necessary number of mea- 20 surements (Western and Bloeschl, 1999;Famiglietti et al., 1999Famiglietti et al., , 2008. Brocca et al. (2007) found that a minimum of 15 to 35 point samples were required for terrain of negligible to significant topography and an extent of around 5000-10000 m 2 . Famigli- The design is split into the selection of (1) an appropriate SMOS pixel, (2) three network clusters within the pixel, and (3) suitable network locations within the clusters. In step 3, a method similar to Friesen et al. (2008) is applied with distribution of the individual stations according to the respective fractions of the prevailing environmental conditions. Friesen et al. (2008) defined the main landscape units a priori, which intro-10 duces a risk to exclude important features from the start. In our much simpler method all environmental information is going into the analysis unchanged, whereupon the most important landscape units of the region are detected. Following this approach, it is anticipated that a priori the likelihood of obtaining a representative large-scale in situ soil moisture average for comparison with SMOS data is strongly enhanced. 15

Study area
The Skjern River Catchment is situated in Western Denmark and covers an area of approximately 2500 km 2 (Fig. 1). The climate in the region is temperate-maritime with winter and summer mean temperatures of around 2 and 16 • C, respectively, and an approximate annual precipitation between 800 to 900 mm. The eastern margin of the 20 catchment is situated at the rim of the ice sheet during the latest glacial advance with mainly loamy soils on undulating calcareous tills. The remaining part comprises the primal fluvioglacial outwash plain consisting of low-relief sandy soils and sediments, while poorly drained basins have been filled with organic material (Greve et al., 2007). The predominant naturally occurring soil type is podsol with a bleached quartz-rich Introduction with often distinct mottling (Scheffer and Schachtschabel, 2002). While water drains quickly through the sandy topsoil, this very firm hardpan is almost water tight causing ponding of water at its surface. When fertilized, limed and irrigated high-yield cultivation is possible; this is the case in the major part of the Skjern River Catchment.
Intermixed are patches of natural vegetation, i.e. grassland, heath and spruce plan-5 tations with pronounced raw humus layers (typically found on podsols). The area is sparsely populated with scattered farms and villages. Within the catchment four study sites were chosen for the HOBE project (Jensen and Illangasekare, 2011, Fig. 1) to assess a wealth of hydrological parameters. The catchment is well-covered with climate and weather stations operated by the Danish 10 Meteorological Institute (DMI). The 24-h precipitation sums presented in this article are extracted from the DMI 10 × 10 km precipitation grid nodes (Scharling, 1999) contained within the SMOS pixel ( Fig. 1). For each day the shelter correction factor of the corresponding month (category B) is applied to the data (Vejen et al., 2000).

Network data
A total of 30 Decagon ECH2O data loggers (Decagon Devices, 2002) were installed, each holding three ECH2O 5TE capacitance sensors measuring soil moisture, temperature, and electrical conductivity (Decagon Devices, 2008) 1 . The 5TE sensors were considered to be a cost-effective solution for large network applications. They 20 are well-suited for measurements in the near-surface layer and they provide integrated measurements over approximately 5-6 cm when installed horizontally (0.3 l measurement volume). Accuracies in mineral soils are ±0.03 and ±1 • C for water content and temperature, respectively. Using the empirical calibration equation of Topp et al. (1980)  volumetric water content is derived from dielectric permittivity, which in turn results from a 5 point dielectric calibration. The TE sensors (predecessors of 5TE) were excessively tested in soils ranging from 3-100 % sand/0-53 % clay and salt-water solutions of electrical conductivities from 1 to 12 dS m −1 by Kizito et al. (2008). They found little probe to probe variability and 5 sufficiently small sensitivity to temperature and electrical conductivity so that one single calibration curve was applicable for all studied conditions. Similarly, for the 5TE sensor type Vasquez and Thomsen (2010) found the Topp equation to be accurate within ±0.02 in the 0-0.5 m depth range at the HOBE agriculture site Voulund (where one network station was placed). Famiglietti et al. (2008) pointed out, that though site-specific calibration is ideal it is impractical for studies with large sensor numbers distributed over a considerable spatial extent. In their 50 km-scale survey they applied a generalized calibration method with an accuracy of ±0.03 to the entire set of probes, and likewise, this was done by Brocca et al. (2010). 15 Given the above findings, the Decagon 5TE calibration equation (Topp et al., 1980) has been applied to the network. The given accuracy has been confirmed by some independent testing (addressed in Sects. 4.3 and 5.1).

SMOS data
The SMOS measurement and soil moisture retrieval concept is complex and will be 20 described to the extent required for understanding the presented work. For further information reference is made to Kerr et al. (2001. The radiation collected by the SMOS radiometer is emitted from the area illuminated by the antenna directional gain pattern (working area, ∼123 × 123 km). Measurements are made at horizontal and vertical polarizations (H and V) and incidence angles rang- 25 ing from around 0 to 60 • as the satellite passes over the terrain. To derive the level 2 (L2) soil moisture product, brightness temperatures T B as acquired by SMOS (proportional to the measured radiation) are modeled for both polarizations at each incidence angle by means of the L-band Microwave Emission of the Biosphere (L-MEB) forward model (Wigneron et al., 2007). An initial soil moisture guess and auxiliary parameters (e.g. soil properties, land cover information, leaf 5 area index, topography, temperature and other climate parameters) are required as input. The soil moisture and temperature initial guesses presented in Sect. 5.3.3 both originate from the European Centre for Medium-range Weather Forecasting (ECMWF) product spatially and temporally aggregated over the working area (both contained in the L2 product). Modeled and measured T B s are compared, and by minimizing a cost function, soil moisture is iteratively retrieved for each node of a fixed earth surface grid (Discrete Global Grid DGG) with uniform spacing (∼15 km). Figure 1 illustrates the locations of the DGG nodes in the Skjern River Catchment, including the working area and corresponding SMOS pixel around one grid node.
L-MEB is based on the relationship between T B , physical temperature and the land 15 surface emissivity/reflectivity, which in turn is related to the soil's dielectric constant after segregating atmosphere, vegetation and surface roughness contributions using the multi-angular and dual-polarized information. Taking advantage of the large contrast between the dielectric properties of water and solid soil particles at L-band, soil moisture is linked to the dielectric constant via the Dobson dielectric mixing model (Dobson 20 et al., 1985;Peplinski et al., 1995). L-MEB is built for uniform scenes with certain model characteristics and calibration parameters. However, the above-mentioned auxiliary input parameters are mostly heterogeneous at significantly smaller spatial scales than SMOS pixels. To account for this, the retrieval algorithm aggregates the estimated contributions from several ele-Introduction Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | radiation fractions originate. A second cluster was allocated to the north-east of the SMOS pixel, around the HOBE agriculture site Voulund with one network station at the study site to render the data connectable to other geophysical measurements. For the same reason one network station was assigned to the HOBE forest site Gludsted, situated some kilometers east of this cluster. The third cluster was placed in the south-5 west to account for the small spatial gradient observed in the mean annual precipitation for the period 1990 to 2005 (10 × 10 km grid, Scharling, 1999) from south-west (∼900 mm yr −1 ) to north-east (∼800 mm yr −1 ).

Selection of theoretical station locations
For positioning the station locations within these cluster areas, a GIS analysis was  Table 1 summarizes soil types with respective grain size distribution and organic matter content of the 0-20 cm topsoil layer (250 m Danish topsoil grid, Greve et al., 20 2007). Accordingly, Table 2 shows the subsoil composition below 30 cm depth (clay versus sand) with corresponding clay contents based on a map from (Bornebusch and Milthers, 1935;Smed, 1979;Schou, 1949;DGU, 1945). In both tables respective soil type fractions contained in the SMOS pixel and working area around node 2 002 029 are given. While the pixel comprises almost 80 % coarse sand in the topsoil and 89 % 25 sand in the subsoil, these percentages are lowered to 46 % and 70 % for the entire working area due to a fractional shift towards more loamy soils concurring with the position of the latest glacial ice margin.   EEA, 2005;Bossard et al., 2000) within the SMOS pixel and working area around node 2 002 029, respectively. They are comparable for the two spatial scales with agriculture taking the major parts, followed by forest (mainly coniferous) and shrub/grassland (heath). In agreement with the corresponding SMOS radiometric fractions, water bod-5 ies only exhibit marginal parts. Land cover exerts strong influence on the SMOS soil moisture algorithm through both choice of the retrieval model and high non-linearity of vegetation parameters. Thus, it is of importance that the area for which the network delivers soil moisture data is representative for the entire working area in terms of land cover, while this is less relevant in case of soil types.
To find the most representative combinations of topsoil, subsoil, and land cover types within the SMOS pixel, the individual data sets were re-sampled and snapped to the land cover 100 m-grid ( Fig. 3a-c). Using the nearest neighbor re-sampling technique merely changed the cell size while all categorical information was conserved. The land cover, top-and subsoil data sets were reclassified to values of 100 s, 10 s and 15 1 digits ("reclass values" in Tables 1-3), and summed up to one grid containing all possible combinations of the original layers (referred to "composite class map" hereafter, Fig. 3d). Figure 4 displays the composite class fractions revealing five classes (212, 232, 412, 512 and 612) with individual shares of >5 %. Together they constitute approximately 75 % of the SMOS pixel and all have a tendency towards very sandy 20 soils. Including the most frequent classes with humus in the topsoil (292) and clay in the subsoil (211), ∼82 % of the prevailing environmental conditions in the validation area are incorporated, which is regarded as a good overall representation. As CORINE land cover class 400 (heterogeneous agriculture) contains all prevailing land cover types (arable land intermixed with forest and shrub/grassland, Bossard et al., winter wheat, and additionally to maize and potatoes (differing plant structure) according to respective fractions.

Field inspection/final decision on station locations
Provisionally, the stations were distributed among the three network clusters using the 10 composite class map (Fig. 3d). Final decisions on the locations were taken after field inspection. Due to an extensive road network, access did not constrain the choice.
For forest and heath (composite classes 512 and 612), no reallocation of the preselected points was necessary, as theoretically estimated land cover and soil types were in good agreement with actual conditions. Three stations were placed under 15 scotch heather, one under natural grass, and four under spruce plantations characterized by pronounced row structure, scarce understory and moss carpets. All these locations exhibit distinct organic surface layers.
The estimated occurrence of agricultural areas and crop types was also encountered in reality, and in case of the composite class 212 the expected very sandy top-and 20 subsoils were clearly perceived at the preselected locations. However, the distinction between classes 212 and 292 (sand and humus in the topsoil, respectively) was almost impossible, as the upper soil layer exhibited a very dark color at all investigated locations, due to intermixed organic matter as a result of agricultural practices. Likewise, for locations where classes with higher clay fractions were indicated on the composite 25 class map (i.e. class 211 with clay in the subsoil or class 232 with loamy sand in the top soil) we could solely notice that soils clearly exhibited greater clay contents than the sandy classes. In situ discrimination between class 211 and 232 turned out to be difficult. Furthermore, at locations where an increased clay fraction was noticed, it persisted usually throughout the entire depth profile. As classes 211, 232 and 292 only account for a small fraction of the entire SMOS pixel (∼13 %), these inaccuracies were accepted when placing the corresponding stations. We resigned the labor-intensive 5 determination of texture and organic amounts for the localization of spots with the exact soil properties inherent in the respective composite classes. Two of the four more clayey stations were placed outside of the third cluster in the south-east to account for the region where the main fraction of more clayey soil conditions within the SMOS pixel occurs as a result of the geomorphological evolution in the area (see Sect. 2).

10
The estimated number of stations per crop type could be maintained, even though some adjustments had to be made between the composite classes ( Table 5). This was accepted since crop rotations change throughout the years. An overview of the final network locations is given in Fig. 2 and Table 6. 15 Sensor installation took place in fall 2009. At each station, three 5TE sensors were placed at respective depths of 2.5, 22.5 and 52.5 cm (corresponding to measurement intervals of ∼0-5, 20-25 and 50-55 cm) from the soil surface after removal of the litter/organic layer (Fig. 5). The sensors were horizontally inserted with the blade in the vertical position to avoid ponding. 20 While for SMOS validation the 0-5 cm data is of most importance, the profile measurements suit the needs of hydrological modeling activities in the HOBE project, possibly in combination with assimilated SMOS data. With respect to heath and forest stations, one 5TE sensor was additionally installed in the organic layer in summer 2010. This is crucial as the signal measured by SMOS over these areas most probably origi-Introduction Sensor readings are logged in 30 min intervals. Stations placed in crops have to be temporarily removed during cultivation practices (seed/plantation and harvest) -twice for summer crops (spring and fall) and once for winter crops (late summer).

Installation
Soil samples were taken at each sensor depth during installation. Sand (2000-20 µm), silt (20-2 µm) and clay (<2 µm) fractions (International Society of Soil Science, ISSS, 1929) of the 0-5 cm depth were determined for all network locations using sedimentation and sieve analysis, and soil bulk density was calculated (Table 6). Additionally, soil samples were collected from 0-5 cm depth on agricultural land, forest and heath (composite classes 212, 512 and 612, resp.) during an airborne campaign (Bircher et al., 2011). These samples were used for calibration checks over the entire 10 wetness range in the laboratory.

Network data analysis
To check the feasibility of our approach as well as the reliability of the network data, several analyses were conducted: The sensor output -sample water content couples from the lab calibration were 15 compared to the Decagon 5TE default calibration curve (Topp et al., 1980). By means of the texture data the actual soil type distribution among the network stations was compared with the one based on the composite class map. Per station the measured soil moisture and temperature data of all depths for the year 2010 was checked for the expected behavior as a function of land cover and soil types. 20 Further network data analyses focused on the 0-5 cm depth only: 1. The soil moisture data of five selected agricultural stations (2.09, 3.08, 3.01, 1.09, and 3.05, Fig. 2 2. To study regional variability and potential influence of the long-term precipitation gradient, soil moisture and temperature of three selected stations of similar texture and land cover in the north-east (1.02, 1.06, 1.09) and south-west part (3.02, 3.04, 3.07) of the SMOS pixel as well as precipitation data of the two closest 10 km grid nodes, respectively, were averaged and compared over the year 2010. 3. Soil moisture and temperature averaged over all 30 network stations were compared with SMOS L2 soil moisture (initial guess and retrieved) and temperature (initial guess) data for the year 2010. Furthermore, to avoid deviations that may arise from the applied petrophysical relationship (Topp et al., 1980), this comparison was also conducted at the dielectric constant level. The 5TE sensor  Fig. 7 the 0-5 cm depth texture data (sand-% vs. clay-%) for the network are shown and compared to the composite classes used in the Danish soil grid (Greve et al., 2007). As the organic content was not measured it is not possible to classify the two stations representing class 292. For the remaining 28 stations it can be seen that: (1) all forest and heath stations (classes 512 and 612) are correctly allocated to the soil type 5 sand, while two of the agriculture class 212 (stations 2.04 and 2.08) exhibit slightly higher clay fractions than expected; (2) the agriculture class 211 is expected only to show more clay conditions in the subsoil, but in fact slightly and significantly higher clay fractions in the topsoil are found for stations 3.01 and 3.08, respectively; (3) with respect to agricultural class 232 station 3.09 is correctly classified whereas station 2.09 10 shows significantly higher clay fractions than expected. Overall, five out of 28 stations are misclassified. However, overall the predetermined number of stations per soil type (Table 4) is more or less maintained in the final network setup. for five selected stations representing the majority of encountered patterns throughout the entire network data set: 2.11 (heath, class 612), 1.04 (forest, class 512), 1.02 (agriculture, class 212, HOBE site Voulund), 2.05 (agriculture, class 212), and 2.09 (agriculture, class 232). Additionally, in case of the heath and forest stations (2.11 and 1.04) data from the sensors installed in the organic layers are depicted. It should be 20 noted that for the organic material, site-specific calibration will be a crucial issue. Thus, at the point of writing this paper these measurements should only be considered in a relative term.

Profile soil moisture and temperature (all depths)
Typically, for all network locations in agricultural fields and with coarse sand in the topsoil, a homogeneous mixture of loose sand and organic material is found in the Introduction surface as a whole. In most cases the 0-5 cm and 20-25 cm sensors were installed in the plow layer. As water infiltrates quickly through the sandy material and the water content in the surface layer is reduced by evapotranspiration, the 0-5 cm sensors generally show drier conditions than the 20-25 cm sensors located just above the hardpan, which restricts the further downward movement of water. Moreover, the 50-55 cm 5 sensors measure high water contents if located close to the upper hardpan boundary (station 1.02) and show much drier conditions when installed within or below the hardpan (station 2.05).
In contrast, a pronounced litter layer of moss/organic material exists (∼5-20 cm) for the sandy soils under natural vegetation. Due to absence of plowing, the topsoil down 10 to the hard pan is leached and quartz-rich as expected for a typical podsol, and the hardpan starts at around 20-25 cm depth. While all four forest stations show similar soil moisture patterns throughout time, the conditions at the four heath stations are very variable. Station 2.11, for instance, is situated in a very wet area where standing water was observable around the station during installation. The 0-5 cm sensor at this 15 station shows high moisture values as it is nourished by the very moist moss/organic layer on top. At the time of installation the 50-55 cm sensor was mounted below the water table. However, when the water table later during the season was lower the effect of the dry sand below the hardpan became evident in the data from the 50-55 cm sensor. In comparison, the sensors in the moss/organic layer as well as the 20 0-5 cm mineral layer of the forest station 1.04 show much drier conditions. This can be attributed to their placement on a small hill. The 20-25 cm sensors ob both stations 2.11 and 1.4 were installed at the upper hardpan boundary and show similar behavior. Generally, the pattern of the forest stations is more related to the one met at agriculture sites where the 50-55 cm sensor is located in the dry sand below the hardpan (station Introduction summer the soil at the 20-25 cm sensor location became drier while the 50-55 cm sensor remained below the water table. The different values for porosity of sandy and clayey soils are well-reflected in the measurements of the 50-55 cm sensors situated below the water table with saturated moisture contents of ∼0.4 and 0.5 in case of the sandy station 2.11 and the clayey 5 station 2.9, respectively. Even higher values are found in the organic material. Furthermore, the effect of texture is also reflected in the seasonal variation of soil moisture for the different soil types. Sandy soils have a smaller water holding capacity compared to clayey and organic materials and as a result the seasonal variation is relatively small. Apart from the drop of moisture content in the organic and 0-5 cm mineral layers dur-10 ing freezing in the winter months this behavior is evident for the two sandy agricultural stations (1.02 and 2.05) as well as for the forest station (1.04). In contrast, the clayey agricultural station (2.09) and the heath station (2.11) show a much higher seasonal variability. Irrigation has obviously a distinct imprint as seen for the agricultural stations 1.02 and 2.05, and in case of the forest site, tree interception must exert a balancing 15 effect.
At all sites the temperature profiles show the expected diurnal and seasonal patterns, as well as a slight time lag and amplitude decrease with increasing depth. Furthermore, the presence of vegetation and moss/organic layers (heath and forest stations) insulating the mineral soil becomes apparent. The isolation effect is reflected in both the 20 diurnal and seasonal temperature amplitudes and is most pronounced for the forest station.
All in all the observed moisture and temperature patterns are clearly related to land cover and soil conditions. Soil moisture seems to be mostly affected by soil characteristics while soil temperature is mostly dependent on land cover. Introduction

Tables Figures
Back Close

Full Screen / Esc
Printer-friendly Version

Interactive Discussion
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | 5.3 Surface soil moisture and temperature (0-5 cm) Figure 9 illustrates the 0-5 cm soil moisture measurements of the agricultural stations 2.09, 3.08, 3.01, 1.09 and 3.05 with similar vegetation and decreasing clay/increasing sand fractions (Table 6), respectively, in comparison with the 0-5 cm average over all 5 30 stations between January and August 2010. The mean of daily precipitation of the 10 km grid nodes contained within the SMOS pixel ( Fig. 1) is also plotted.

Texture comparison
Over the major part of the chosen time span, increasing clay content complies with higher moisture content, resulting in significant overrepresentation with respect to the overall network average, and vice versa in the case of high sand contents. Thus, 10 the influence of soil texture is clearly demonstrated and also reflected in the biases (average residuals from expected value) ranging from 0.146 for the clay station 2.09 to −0.057 for station 3.05 with highest sand fractions. The larger absolute bias (relative to other stations) of the clay station is reasonable, as the 30 station average contains a much larger fraction of sandy sites. The moisture pattern also follows the precipitation 15 trend well, and in March, snow melt is observable throughout all stations. Figure 10 shows average and standard deviation (shaded region) of the 0-5 cm soil moisture and temperature of three selected stations of similar texture and land cover in the north-east (1.02, 1.06, 1.09) and south-west part (3.02, 3.04, 3.07) of the SMOS 20 pixel, as well as 24 h precipitation accumulations of the two closest 10 km grid nodes, respectively, for the year 2010.

Regional comparison
Regional differences are most pronounced for soil moisture and least for temperature. However, in any case they are small with low RMSE/biases ( 0.99 and 0.86. Moreover, with temporal mean standard deviations of 0.024 and 0.041 (soil moisture) and 0.37 and 0.5 • C (temperature) for the north-eastern and southwestern stations, respectively, the variability between the two areas is in the same order as within them.

SMOS L2 comparison
5 Figure 11 displays 0-5 cm average network and SMOS soil moisture and temperature data (L2 product) for the year 2010, as well as the corresponding mean of daily precipitation of the DMI 10 km grid nodes contained within the SMOS pixel (Fig. 1). Also network soil moisture spatial variability (standard deviation, blue shaded region) and in situ sensor accuracy are shown (grey shaded region). For SMOS retrieved val-10 ues and associated radiometric accuracies (shaded region) as well as the initial guess are shown. Mean network soil moisture fluctuates around a temporal average of 0.176 and with a standard deviation of 0.041. The spatial variability between the individual stations is larger with a temporal average of 0.070, which is in the same order as found by Famiglietti et al. (2008) for a site in the United States at the same spatial scale. 15 Network and SMOS soil moisture follow the precipitation dynamics well. Correlations (R 2 ) between network and SMOS retrieved and initial guess soil moisture, respectively are 0.49 and 0.67. However, remarkable offsets are visible. While the SMOS soil moisture initial guess approximately corresponds to the upper boundary of the network variability error bar, the retrieved data follows more or less its lower boundary, or even relations, respectively. Consequently, there is no distinct difference between the two dielectric models with R 2 s equal to that for the soil moisture comparison. As the SMOS dielectric constant is computed from retrieved soil moisture by means of the Dobson model, this implies that at both comparison levels the uncertainty is consistent and remains on either the network or the SMOS data side. 15 Based on the results of the presented network data analyses together with the fact that our findings from the comparison with SMOS data are well in range with worldwide validation results, we consider the network to operate according to expectations and to be well-suited for SMOS validation. The discrepancies between network and retrieved SMOS soil moisture data need to be more closely investigated. Currently, numerous 20 explanations are under discussion: (1) a mismatch between sampling depth of conventional soil moisture sensors (∼5-7 cm) and the depth contributing to L-band soil emission (<5 cm, Escorihuela et al., 2010), (2) scale effects due to the large disparity in spatial scale between the SMOS and in situ measurements, (3) inaccuracies in the SMOS retrieval algorithm and related input, (4) inacurracies in the in situ measure- Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | means of the carefully chosen network setup. Meanwhile, we see RFI contamination and inaccuracies in the SMOS retrieval algorithm as probable causes for the bias. Currently, the replacement of the Dobson dielectric mixing model with the one of Mironov (Mironov et al., 2004) is for example under investigation. Bircher et al. (2011) showed that Mironov performed better at the Danish validation site to bring brightness temper-5 atures modeled from in situ soil moisture data in agreement with airborne brightness temperature measurements at the 2 × 2 km scale. Thus, it is also likely that the deviances between SMOS and in situ soil moisture could be lowered by using Mironov in the SMOS retrieval algorithm. With respect to the high amplitudes in the retrieved SMOS data, there is generally consensus that they are likely to be attributed to the 10 mismatch in sampling depth. Generally, the very top layer shows a rapid soil moisture increase immediately following rain events, succeeded by a fast decrease as a result of evaporation and infiltration processes. At deeper depths this response is delayed and somewhat less. The wetter and the more sandy the soils, the more pronounced this effect is. However, at this point, this remains a hypothesis. Further investigations are 15 needed to separate the respective contributions to the deviations between in situ data and SMOS and thus clarify these issues.

Conlusions
A soil moisture and temperature network with 30 stations (sensors at 0-5, 20-25 and 50-55 cm depths plus in the organic layer in the case of heath/forest locations) has 20 been established within one SMOS pixel (44 × 44 km) in the Skjern River Catchment, Western Denmark The design of the network included the following phases: (1) the selection of SMOS pixel 2 002 029 with minimal water fraction and maximal catchment coverage, (2) the arrangement of three network clusters along a long-term precipitation gradient centered at the SMOS node, and (3)  conditions. In case of agriculture, additionally crop type frequency was considered. Using this method, it was possible to obtain a representative large-scale in situ soil moisture average for comparison with SMOS data. Analysis of the collected network data during the year 2010 showed that soil moisture generally follows the precipitation trend. Furthermore, soil moisture and tempera-5 ture patterns were relatable to the respective land cover and soil conditions. The high soil moisture variability throughout the stations seems to be a strong function of texture/structure while to a less extent influenced by land cover. At the same time the variability in soil temperature is less pronounced and merely a function of the latter. Regional differences in 0-5 cm soil moisture, temperature and precipitation between 10 the north-east and south-west turned out to be small.
A first comparison between 0-5 cm network averages and the SMOS L2 product showed comparable trends with R 2 of 0.49/ 0.67 and 0.97 for SMOS retrieved/initial soil moisture and initial temperature, respectively. The two former indicate significant under-/overrepresentation of the network data (biases of −0.092/0.057 m 3 m −3 ) as well 15 as faster and stronger wetting/dry-downs (larger amplitudes). Correlation with precipitation is traceable in both, network and SMOS soil moisture data. Average network and SMOS soil temperatures are in good agreement with a bias of −0.2 • C. Thus, this parameter should not introduce errors in the soil moisture retrieval process. Based on these findings together with the fact that our SMOS data comparison is 20 well in range with worldwide validation results, we consider the network to operate according to expectations and to be suitable for SMOS validation. Extensive validation activities are currently ongoing at the Danish validation site. It is likely that the discrepancies between network and SMOS soil moisture result from a combination of several factors. The investigation of these potential error sources and their respective contribu-Introduction Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Brocca, L., Morbidelli, R., Melone, F., and Moramarco, T.: Soil moisture spatial variability in experimental areas of central Italy, J. Hydrol., 333, 356-373, 2007. 9964 Brocca, L., Melone, F., Moramarco, T., and Morbidelli, R.: Spatial-temporal Devices, Inc., 2365NE Hopkins Court Pullman WA 99163 USA, version 3 Edn., 2008. 9966 DGU: Geologiske kort over Danmark, 1:100.000, D.G.U, 1945 Microwave dielectric behavior of wet soil -Part II: Dielectric mixing models, IEEE T. Geosci. Remote Sens., 23, 35-46, 1985. 9968 5 Dorigo, W. A., Wagner, W., Hohensinn, R., Hahn, S., Paulik, C., Drusch, M., Mecklenburg, S., van Oevelen, P., Robock, A., and Jackson, T.: The International Soil Moisture Network: a data hosting facility for global in situ soil moisture measurements, Hydrol. Earth Syst. Sci. Discuss., 8, 1609-1663, doi:10.5194/hessd-8-1609  Hydrotope-based protocol to determine average soil moisture over large areas for satellite calibration and validation with results from an observation campaign in the Volta Basin, West Africa, IEEE T. Geosci. Remote Sens., 46,[1995][1996][1997][1998][1999][2000][2001][2002][2003][2004]2008