The Maqu soil moisture monitoring network was set up in July 2008 on the north-eastern fringe of the Tibetan Plateau (33°30′–34°15′N, 101°38′–102°45′E), located in the southern part of Maqu county, on the border between Gansu and Sichuan provinces in China. A detailed description of the network is reported in Dente et al., [6] and in Su et al., [7]. The network is located at the first major bend in the Yellow River, where the landscape is characterised by the large river valley and surrounding hills with an elevation ranging from 3,200 m to 4,200 m a.s.l. The Maqu region is shown in Figure 1 by a Landsat 5TM image collected in September 2007, with the locations of the monitored sites highlighted as white rectangles. The land cover consists of uniform short grassland with silt loam soils and the wetlands cover a large part of the valley. The climate is characterised by dry and cold winters (November–March) and by a rainy and relatively warmer monsoon season (April–October). Soil moisture and soil temperature are continuously measured by means of EC-TM ECH₂O probes (Decagon Devices, Inc., USA) at a depth of 5 cm (and in deeper layers) at 20 sites every 15 min. The monitoring sites are distributed over an area of approximately 40 km × 80 km and are characterised by a variety of altitudes and slopes and differing soil characteristics. The network set up ensures that the soil moisture spatial variability of the Maqu region is well monitored, and the spatial average of the measurements collected at each site can be considered an accurate indicator of the soil moisture dynamics at the network scale, as shown in [6].

A detailed land cover map is available for the Maqu region, the Multi-source Integrated Chinese Land Cover (MICLCover) map provided by the Cold and Arid Regions Environmental and Engineering Research Institute of the Chinese Academy of Sciences (CAREERI/CAS). The MICLCover map was obtained by integrating several classification maps over China including a vegetation map, a land use map for the year of 2000, a swamp-wetland map, a glacier map and a Moderate-Resolution Imaging Spectroradiometer land cover map for 2001 (MODIS2001) using a decision-fuse method based on the Dempster-Shafer evidence theory [8]. The accuracy of the MICLCover map was assessed at 71% all over China, which was shown to be higher than the MODIS2001 map accuracy.

The Twente soil moisture monitoring network is located in the eastern part of the Overijssel province in The Netherlands, mainly covering the region called Twente, but also part of the Salland region and the Gelderland province (52°05′–52°27′N, 6°05′–7°00′E). The region is flat, with an elevation ranging between 3 m to 50 m a.s.l. The most extensively occurring land cover is grassland for pasture which is harvested and fertilized several times a year. However, the land use of this region also includes a mosaic of agricultural fields, forest patches and several urban areas. A Landsat 5TM image collected in June 2010 of the Twente region is shown in Figure 2, with the locations of the monitored sites highlighted as white rectangles. The main crop is corn, which is planted in April and harvested in September. In this region, the soil texture of the surface layer is mainly sand and loamy sand. Soil moisture and soil temperature have been continuously monitored since July 2009 at 20 sites, spread across an area of approximately 50 km × 40 km at a depth of 5 cm every 15 min. The sites were selected in order to monitor the area extensively for a variety of soil types and land covers. However, one of the main factors affecting the spatial variability of soil moisture in this area, besides land cover, is the ground water table. Due to groundwater management occurring in this region, the groundwater table may rise to 25 cm below the soil surface as well as drop to below 160 cm. EC-TM ECH₂O probes (Decagon Devices, Inc., USA) are used to measure the soil moisture and temperature at 5 cm depth, as well as in deeper layers. A detailed description of the network and the dataset is reported in Dente et al., [9].

SMOS Level 2 (L2) soil moisture products (reprocessed data, version 5.05 of the Level 1 processor and version 5.01 of the soil moisture Level 2 processor) from January 15 to December 15, 2010 were used in this study. The products are geolocated on an equal-area grid called Icosahedral Snyder Equal Area (ISEA), where the inter-cell distance is uniformly 15 km. The equator crossing time of ascending and descending passes is at 6 a.m. and 6 p.m. (local solar time), respectively, and the complete globe is covered every 3 days.

An iterative approach is followed in order to retrieve soil moisture and vegetation optical thickness. This approach is based on the principle of finding the optimal set of soil moisture and vegetation parameters that minimizes the difference between modeled and measured brightness temperature [5] for a variety of incidence angles, for a given surface cover and soil texture. The obtained soil moisture is an estimate for the surface layer of the soil, a few centimeter deep [10]. The retrieval algorithm is described in detail in the Algorithm Theoretical Basis Document [11]. The forward model used in the SMOS Level 2 processor is the L-band microwave emission of the biosphere (L-MEB) model [12]. L-MEB is based on a simplified zero-order radiative transfer equation [13], where the rough soil surface contribution to the brightness temperature is modeled following the approach by Wang and Choudhury [14] and the vegetation contribution is modeled following the τ − ω approach, i.e., the optical depth τ and the single scattering albedo ω are used to parameterize the vegetation attenuation and the scattering effects within the canopy layer. The soil effective temperature is computed as a function of soil temperature at the soil surface and at increasing depth, as described in Wigneron et al., [15]. The soil dielectric mixing model by Dobson et al., [16] is used to relate the soil permittivity to soil parameters such as soil moisture.

The retrieval is carried out at each node of the ISEA grid and the parameterization of L-MEB depends on the dominant land cover at the node and its surrounding area. Information on the land cover class along with soil and vegetation properties are provided to the SMOS L2 processor by the ECOCLIMAP land surface global database [17] at a resolution of 1 km.

The land cover map in the ECOCLIMAP database is obtained by integrating various data sources, including two global land cover datasets: the International Geosphere-Biosphere Programme Data and Information System (IGBP-DIS) [18,19] and the University of Maryland database (UMd) [20]. The location of the surfaces permanently covered by ice and wetlands is provided by the IGBP-DIS database, whereas all the other cover types are derived from the UMd global land cover map.

In the L2 processor working area of 123 km × 123 km surrounding each node (the distance of 123 km corresponds to twice the largest extent of the 3 dB footprint occurring in SMOS soil moisture observations), different land cover types may be present, contributing to the surface microwave emission. Therefore, the weight of each contribution is determined taking into account the cover fraction and the SMOS radiometric sensitivity in the antenna pattern. In this way, the dominant cover type is assessed and if it consists of a surface type where soil moisture estimation is meaningful (for example low vegetation or forest), retrieval is carried out. For the non-dominant cover type of the node, the contribution to the overall emissivity of the node is taken into account by the L-MEB model, but no retrieval is performed. Any given node is covered by several views, each having a different spatial extent. Therefore, the application of the weight function may give different results for each view.

Auxiliary data, such as soil temperature at the surface and at depth, initial soil moisture values and information about snow cover, are provided by the European Centre for Medium-Range Weather Forecasts (ECMWF) model. When the ECMWF surface soil temperature drops below a given threshold, the soil is considered frozen and soil moisture data are not retrieved. When, according to the ECMWF model, dry snow is present, then the snow is considered transparent and retrieval is carried out as if there was no snow.

The L2 soil moisture products contain not only the retrieved parameters (soil moisture and optical thickness), but also ancillary information obtained during the retrieval, such as surface temperature, roughness parameters and dielectric constant, as well as several confidence, processing and science flags. The confidence and processing flags refer to the quality and characteristics of the retrieval, retrieval options and conditions adopted during the processing and information on L1c data used for the retrieval. The flags useful for the analysis carried out in this study include those indicating: if the retrieval values were outside the expected range; if there was a poor fit between modeled and observed brightness temperature; if the retrieval error was above a certain threshold. The science flags refer to features of the surface that might affect retrieval, for example presence of snow, frozen soils, wetlands and forest. Information about the suspected presence of RFI can be obtained from two bands of SMOS L2 products, N_RFI_X and N_RFI_Y, reporting, for any given node, the number of brightness temperature measurements in X and Y polarization, respectively, disregarded in the retrieval process because of suspected RFI. Combining this information with the total number of brightness temperature measurements available for the given node (M_AVA0) provides the percentage of brightness temperature measurements where RFI is suspected.

The time series of SMOS data spatially averaged over the area of interest and the time series of the in situ average were compared, as shown in Figure 3 for the Maqu region and in Figure 4 for the Twente region. The comparison was done separately for ascending and descending products.

The seasonal behavior of SMOS L2 products is generally in agreement with the soil moisture behavior for both the Maqu and the Twente region. SMOS data on average show relatively low values in the Maqu region in winter with an increase in the monsoon season, following the in situ behavior. Instead, in the Twente region the in situ data show higher soil moisture values in winter and a decrease in summer with the SMOS data depicting a similar pattern.

However, noise was found in the satellite-derived products, consisting of variations in the SMOS estimates not corresponding to in situ data variations, and affecting the agreement between the two datasets. In both regions, the descending products are noisier than the ascending ones. Contribution to the noise is certainly partly due to the fact that several SMOS products covered only partially the area of interest, where the soil moisture retrieval in the rest of the area was unsuccessful. This might affect the validation results, as the SMOS data covering part of the area were compared with the spatial average of the complete network. The empty symbols in Figure 3 and 4 indicate when SMOS products were not generated for at least half of the area, whereas filled symbols refer to a good coverage. The figures show that several SMOS products only partially covered the Maqu region, as well as the Twente region. In general, the two figures show that retrieval was more problematic in the Maqu region than in the Twente region, as data are missing more frequently over the Chinese site with more gaps in the time series, than over the Dutch site. The reason for data in space and time having been missed is investigated in Section 4.2. However, the agreement between the variations found in SMOS data and in situ data does not improve by leaving out the products with partial coverage of the area of interest.

From the time series comparison one can also conclude that the SMOS soil moisture is affected by general underestimation, which appears to be larger for the ascending products over both Maqu and Twente, than for the descending ones. Though with a bias, the SMOS data fall within the variability range of the in situ data for the Maqu region. However, for the Twente region several SMOS data are outside the in situ variability range.

The SMOS mission commissioning phase ended in May 2010, therefore it might be that the data of the first few months of the year have a low quality. However, the agreement between satellite-derived and in situ data does not show any clear dependency on this factor. The underestimation, as well as the noise, does not occur in a specific period of the year or for a particular soil moisture variability range.

The reasons for the noise and underestimation in the SMOS soil moisture products, as well as for the data missing in space and time, are investigated in the following sections. Confidence and science flags were analyzed to obtain information about retrieval algorithm characteristics and problems occurring in the retrieval process.

The reason of an unsuccessful retrieval over part of or the whole area of interest was investigated by checking the flags which indicate the supposed presence of frozen soil and RFI and checking the flags which indicate the performance of the retrieval algorithm, such as the flag for out-of-range retrieved values and the flag for high retrieval error. Therefore, in this section only the SMOS nodes without retrieval are included in the analysis.

As mentioned in Section 2.3, no retrieval is performed when the soil is frozen. Figure 5 (top-left) shows the percentage of nodes included in the Maqu region with a raised frost flag over the total number of nodes in the region lacking retrieval, for both the case of no retrieval over a part of the area and the case of no retrieval over the whole area. The same plot is presented for the Twente region in Figure 6. Figure 5 shows that the complete area of the Maqu network was characterized by the frost flag several times in winter, resulting in no data, whereas the Twente region was not considered to be covered by frozen soil except for a few days in January and December. When the soil moisture was only retrieved over part of the area, frozen soil was not the main cause of unsuccessful retrieval for the rest of the area.

However, the analysis revealed that for several products of the winter season when no soil moisture was retrieved, the Maqu region was completely frost flag free or only partially covered. The absence of the flag for frozen soils in winter over the whole or part of the Maqu area is in disagreement with the in situ measurements. The spatial average of the soil temperature measured in situ at 5 cm depth in the Maqu region is shown at the bottom left of Figure 5. The measurements at the Maqu site show that the soil temperature at 5 cm depth dropped below zero from January to February and from November to December 2010 at all sites of the network, meaning the upper layer of the soil was homogenously frozen in winter, whereas in March and October the frost conditions might not be spatially homogenous, though still present. The soil froze several times in January and February in the Twente region as well, as shown in Figure 6, but the frost flag was raised only once in the SMOS data. As mentioned in Section 2.3, the frozen soil flag is raised when the surface temperature falls below a certain threshold, input that is provided by the ECMWF model. Therefore, it is possible that surface temperatures provided to the SMOS L2 processor were not accurate, but further analysis would be needed to investigate this issue.

The main reason for unsuccessful retrieval, in particular when there is a partial SMOS data coverage over the area of interest, was found to be related to retrieval algorithm performance. Indeed, for most of the nodes where no data were retrieved, the confidence flags indicated that the retrieval attempt was unsuccessful because the retrieved parameters were out of range or the error of retrieval exceeded the maximum threshold (see Figures 5 and 6). However, it was not possible to find further information in the L2 products as to what exactly caused this and at which level of the retrieval processing the problem occurred.

The suspected presence of RFI might be another reason for disregarding some SMOS data in the retrieval process, therefore the SMOS L2 product bands related to RFI were analyzed. The number of disregarded SMOS L1c data per node because of suspected RFI, averaged over the Maqu and Twente region, and expressed as a percentage of the total number of brightness temperature observations (obtained as described in Section 3), is plotted in Figure 5 for Maqu and Figure 6 for Twente. The presence of RFI is definitely much higher in the Twente region than in the Maqu region. At the Dutch site it occurred several times that almost all the SMOS observations were affected by RFI, in particular when no retrieval was attempted over the entire network area. This was not the case for the Maqu region, where the site was suspected to be completely covered by RFI for only a few days. However, the Maqu region was not free of RFI on other days. Indeed, an average of 30% of SMOS observations over the nodes without retrieval was considered to be affected by this interference.

The SMOS soil moisture data were further analyzed, in order to investigate the causes of noise and inaccuracy. Therefore, different flags and bands of the SMOS products were checked for all nodes where soil moisture was retrieved, allowing understanding of which input information was used in the retrieval process that might have caused the noise.

Figures 7 and 8 show that in all the data retrieved over the Maqu region and in most of the data retrieved over the Twente region the flag indicating a poor fit between modeled brightness temperature and SMOS observations was raised. This revealed that the performance of the retrieval algorithm were not optimal over either sites, probably due to inaccurate inputs.

As already mentioned in Section 4.2, the surface temperature used as input to the retrieval algorithm was very probably not accurate, in particular in winter, implying an inaccurate detection of frozen soils. According to the in situ measurements (Figures 5 and 6), the soil surface was homogenously frozen in January and February and in November and December in the Maqu region, and in February for several days in the Twente region. Therefore, no SMOS soil moisture data were expected in these periods. However, several data are available, as the frozen soil was not correctly detected. Dry snow was sometimes assumed to be present (Figures 7 and 8), but retrieval was performed as in the nominal conditions, as the snow was considered transparent (Section 2.3).

Frozen soils and snow might have caused an inaccurate retrieval not only in the mentioned periods, but also in the periods shortly before and shortly after, as the region of interest could have been partially frozen or partially covered by snow.

In order to have a clearer view of the match between retrieved and measured soil moisture, the SMOS soil moisture spatial average was compared with the in situ measurements collected at the time of the satellite acquisitions, as shown in the scatter plots in Figure 12. Moreover, Table 1 reports the determination coefficient, R², and the root mean square error, rmse, between the two time series. The analysis was done separately for ascending and descending data.

For both the Maqu and the Twente region, there is closer agreement of the SMOS soil moisture retrieved from ascending data and the in situ data, than from descending data, though R² remains relatively low, i.e., 0.55 for the Maqu region and 0.51 for the Twente region. The plots show that the dry bias in the SMOS retrieval is more evident in the ascending case than in the descending case, but that the data are more spread out in the latter case than in the former. A quantification of the dry bias of the SMOS data was obtained with the computation of the rmse. Products retrieved from ascending data over the Twente site show on average a dry bias of approximately 0.17 m³/m³, whereas those over the Maqu site show a bias of approximately 0.13 m³/m³. The correlation and the dry bias only slightly improves, if the SMOS data covering less than half the area of interest are disregarded (empty symbols in Figure 12), though not for the descending data over Maqu. Disregarding the periods of possible presence of frozen soils and snow does not improve the match between the two datasets over Maqu and only slightly increase the determination coefficient over Twente. Moreover, in order to check if the presence of mountainous areas, the river and the cities as well as the presence (or assumed presence) of forest in the observed area affected the agreement between satellite-derived and in situ data, the R² and rmse were also computed, disregarding the nodes covering these targets in the SMOS spatial average. Table 1 shows that the R² and rmse does not significantly change when the mountainous areas and the river of the Maqu region and the largest cities of the Twente region are excluded by the computation. Moreover, R² and rmse does not improve when the nodes with a raised forest flag are excluded from the SMOS spatial average. This leads to the conclusion that noise and bias of the SMOS soil moisture data over Maqu and Twente region are not due to the presence of these targets and are not related to the inclusion in the L-MEB model of a forest contribution. Moreover, they are not limited to the period of frost and snow.

The difference between the depth of the soil layer contributing to the brightness temperature measured by SMOS radiometer and the soil depth where the measurements were carried out certainly affects the relationship between satellite-derived and in situ data. Measuring soil moisture at deeper layers than the penetrating depth of the microwave radiation should lead to an underestimation in the dry season (as the upper soil layers are dryer than the deeper layers) and an overestimation in the rainy season or soon after a rainy event (as the upper layers are wetter than the deeper ones), but this is not the case for the SMOS data, as the dry bias is quite systematic in all seasons. Therefore, this effect does not explain the SMOS bias.

The rmse found over the Maqu and Twente region are much higher than the SMOS mission requirements for soil moisture observations, i.e., 0.04 m³/m³ [2]. The dry bias and noise in the SMOS data are not specific for the Maqu and the Twente sites, as they were also found at the upper Danube catchment (southern Germany) [23], at two sites in Italy and one in Luxemburg [24], at the REMEDHUS network area in Spain [25] and at several sites of the SCAN network in the U.S. [26]. However, there are also sites where bias was almost completely absent, as it was shown at the Walnut Gulch (AZ) watershed, the Little Washita (OK) watershed and the Little River (GA) watershed in the U.S. [27] as well as at some of the sites of the SCAN network [26]. Considering all these studies one could conclude that the dry bias is not dependent on a specific land cover, surface condition or season, though it might be related to the presence of RFI. However further analyses are needed to understand the cause.