Evaluation of soil moisture downscaling using a simple thermal-based proxy – the REMEDHUS network ( Spain ) example

Soil moisture retrieved from satellite microwave remote sensing normally has spatial resolution on the order of tens of kilometers, which are too coarse for many regional hydrological applications such as agriculture monitoring and drought prediction. Therefore, various downscaling methods have been proposed to enhance the spatial resolution of satellite soil moisture products. The aim of this study is to investigate the validity and robustness of the simple vegetation temperature condition index (VTCI) downscaling scheme over a dense soil moisture observational network (REMEDHUS) in Spain. First, the optimized VTCI was determined through sensitivity analyses of VTCI to surface temperature, vegetation index, cloud, topography, and land cover heterogeneity, using data from Moderate Resolution Imaging Spectroradiometer (MODIS) and MSG SEVIRI (METEOSAT Second Generation – Spinning Enhanced Visible and Infrared Imager). Then the downscaling scheme was applied to improve the spatial resolution of the European Space Agency’s Water Cycle Multi-mission Observation Strategy and Climate Change Initiative (ESA CCI) soil moisture, which is a merged product based on both active and passive microwave observations. The results from direct validation against soil moisture observations, spatial pattern comparison, as well as seasonal and land use analyses show that the downscaling method can significantly improve the spatial details of CCI soil moisture while maintaining the accuracy of CCI soil moisture. The accuracy level is comparable to other downscaling methods that were also validated against the REMEDHUS network. Furthermore, slightly better performance of MSG SEVIRI over MODIS was observed, which suggests the high potential of applying a geostationary satellite for downscaling soil moisture in the future. Overall, considering the simplicity, limited data requirements and comparable accuracy level to other complex methods, the VTCI downscaling method can facilitate relevant hydrological applications that require high spatial and temporal resolution soil moisture.

olution of satellite soil moisture products. The aim of this study is to investigate the validity and robustness of the simple Vegetation Temperature Condition Index (VTCI) downscaling scheme over a dense soil moisture observational network (REMEDHUS) in Spain. Firstly, the optimized VTCI was determined through sensitivity analyses of VTCI to surface temperature, vegetation index, cloud, topography and land cover het-

Introduction
Soil moisture (SM) is known to be an important state variable that determines the partitioning of surface net energy into latent and sensible heat fluxes, as well as the partitioning of precipitation into infiltration and runoff (Porporato et al., 2004;Vereecken et al., 2014). In the context of global climate change, accurate information of soil 5 moisture is of great importance for advancing our understanding of the energy and mass exchanges between the atmosphere, hydrosphere and biosphere (Petropoulos et al., 2015;Seneviratne et al., 2010). In addition, soil moisture is important for numerous practical applications such as irrigation water management (Bastiaanssen et al., 2000), ecological modeling (Nemani et al., 2009), vegetation productivity estimation that (1) dielectric constant measurement can be related directly to soil moisture; (2) soil moisture can be retrieved regardless of the atmospheric conditions (Hain et al., 2011;Loew et al., 2006). To date, several global microwave based soil moisture products are available, such as the ASCAT soil moisture product (Wagner et al., 1999;Naeimi et al., 2009), the AMSR-E and AMSR2 soil moisture products (Owe et al., 2008;Pari-5 nussa et al., 2014), the SMOS soil moisture product (Kerr et al., 2001;Jacquette et al., 2010), as well as the European Space Agency's Water Cycle Multi-mission Observation Strategy and Climate Change Initiative (ESA CCI) soil moisture product. The CCI SM product is a merged product based on six microwave products (Liu et al., 2012(Liu et al., , 2011Wagner et al., 2012). These soil moisture products normally have spatial resolu- 10 tion on the order of tens on kilometers, which serves well for global scale applications. However, this spatial resolution is often too coarse for regional and local applications such as agriculture monitoring and drought predication, which normally require a spatial resolution of 1-10 km (Crow et al., 2000;Piles et al., 2011).
Optical/thermal infrared (TIR) sensors can provide complementary information of 15 soil moisture patterns at higher spatial resolutions (tens of meters to several kilometers). The surface reflectance observed by optical sensors can be used to explore the state of soil moisture indirectly through empirical spectral vegetation index (Gao et al., 2013;Lobell and Asner, 2002 Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | ted, the shape of the scatterplot generally resembles a physically meaningful triangular or trapezoidal feature space, due to different sensitivity of surface temperature to soil moisture variations over bare soil and vegetation covered areas (Carlson et al., 1994;Peng et al., 2013b). Based on this feature space, several indexes such as Vegetation Temperature Condition Index (VTCI) (Wan et al., 2004) and Temperature Vegetation 5 Dryness Index (TVDI) (Sandholt et al., 2002) have been widely used for assessing the status of soil moisture and monitoring drought condition (Patel et al., 2008;Karnieli et al., 2010;Mallick et al., 2009;Peng et al., 2013a). Similarly, Chauhan et al. (2003) proposed a soil moisture downscaling scheme that links the soil moisture with surface temperature, vegetation index and surface albedo through linear regression equation. 10 Following this idea, some other studies have tried to improve the regression models by including other inputs such as brightness temperature and surface emissivity (Piles et al., 2014;Sobrino et al., 2012). Recently, Peng et al. (2015) proposed a new and simple method to improve the spatial resolution of microwave soil moisture with VTCI as the unique downscaling factor. They demonstrated the feasibility of the proposed 15 method via validation against limited ground-based soil moisture measurements and spatial comparison with land cover map. However, to further investigate the robustness of the proposed method, Peng et al. (2015) suggested that more validation work against dense soil moisture observational networks is required. Therefore, the present study focuses mainly on investigating the validity and robust-20 ness of this simple downscaling scheme through comparison with ground-based soil moisture measurements from a dense observational network (REMEDHUS) in Spain (Martínez-Fernández and Ceballos, 2003). The REMEDHUS site has already been widely used for validation of soil moisture estimates from remote sensing (Brocca et al., 2011;Ceballos et al., 2005;Saìnchez et al., 2012). This study has two major objectives. 25 First it explores and analyzes the sensitivity and robustness of VTCI on land surface temperature, vegetation index, clouds, terrain condition and heterogeneity of land cover and validates the accuracy of soil moisture downscaling using VTCI over the REMED-HUS site. Secondly it investigates the merit of using of geostationary satellite data for downscaling soil moisture. The latter is done, as polar orbiting satellites like e.g. the as Moderate Resolution Imaging Spectroradiometer (MODIS) and Advanced Very High Resolution Radiometer (AVHRR) are in general used for downscaling microwave soil moisture, while the geostationary satellite data are rarely applied. As geostationary observations can provide more cloud free observations due to their high temporal res-5 olution, they have the potential of estimating the thermal inertia at higher frequencies (Fensholt et al., 2007;Shu et al., 2011;Stisen et al., 2008). Thus, the results from both, polar orbit satellite (MODIS Terra/Aqua) as well as geostationary satellite (MSG SEVIRI) were used in the current study to downscale ESA CCI soil moisture product. To the best of our knowledge, this is also the first study to inter compare the performances 10 of geostationary and orbit satellites for downscaling soil moisture.

Study area and REMEDHUS observation network
The current study is carried out in Spain, where a central area is selected for downscaling CCI SM due to its relatively flat characteristic (Fig. 1). The land cover of this region is dominated by croplands and shrublands, and the elevation of most of the area 15 is larger than 450 m a.s.l. (Zhang et al., 2014). The region has a continental semiarid Mediterranean climate, which is characterized by dry and warm summers and cool to mild and wet winters (Castro et al., 2004). The REMEDHUS soil moisture observation network is in the central part of the study area and shown in Fig. 1  in Table 1. The REMEDHUS network has been widely used for different applications such as parameterization of water balance models (Sánchez et al., 2010), calibration and validation of soil moisture product from remote sensing (Saìnchez et al., 2012;Wagner et al., 2008), and especially evaluation of downscaled soil moisture product from SMOS and AMSR-E (Piles et al., 2014;Sánchez-Ruiz et al., 2014;Zhao and Li, 5 2013). These studies presented more complex downscaling approaches than the one explored in this study. They provide a baseline to cross compare the downscaling results of this study with their results. Therefore, study period from 1 January 2010 to 31 December 2011 is chosen to be similar to the published studies and to make the inter-comparison more reasonable. The soil moisture measurements are obtained from 10 the International Soil Moisture Network (ISMN, https://ismn.geo.tuwien.ac.at) (Dorigo et al., , 2013.

Satellite data
Several satellite platforms with different temporal and spatial resolutions are used in this study. They provide different land surface products such as soil moisture, land sur- 15 face temperature and vegetation indexes. Table 2 gives an overview of these satellite products used in our study.

ESA CCI soil moisture
The ESA CCI soil moisture is a unique multi-decadal (35 years from 1978 to 2013) satellite-based soil moisture dataset on a daily basis and at a spatial resolution of 20 0.25 • , which has great potential for climate related studies and applications (Loew et al., 2013 gether with a cumulative distribution function (CDF) matching technique. The detailed harmonization procedure is described in Liu et al. (2012Liu et al. ( , 2011. The first version of the CCI SM dataset (01.0) was published in 2012 covering the 32 years period from 1978 to 2010. To extend the temporal coverage to 2013, improve the gapping filling and processing algorithm, a new version of CCI SM product (02.1) was released recently. It contains three soil moisture products: active only, passive only, and merged active-passive product. A comprehensive validation of the merged active-passive product using 596 sites from 28 different observation networks worldwide was carried out by Dorigo et al. (2014). It was shown that the CCI SM product has a mean correlation coefficient (R) of 0.46 and an average unbiased root mean square deviation (ubRMSD) 10 of 0.05 m 3 m −3 on daily timescales. Similarly, Zeng et al. (2015) provided an evaluation of seven popular satellite-based and one reanalysis soil moisture products over the Tibetan Plateau using in-situ observations. They concluded that the CCI SM product correlates better with in-situ observations with the highest R and the smallest ubRMSD compared to other products. Furthermore, the trend of CCI SM highly agrees with the 15 trends of precipitation, and vegetation vigor from various reanalysis products (Albergel et al., 2012;Dorigo et al., 2012). Therefore, the CCI merged soil moisture is selected in our study to explore the possibility and potential of downscaling to high spatial resolution with optical/thermal infrared data. soil moisture (e.g. Srivastava et al., 2013;Choi and Hur, 2012). In this study, both MODIS/Terra and MODIS/Aqua are used to downscale soil moisture. The MODIS products used in this study are Collection 5 MODIS land surface temperature and vegetation indexes (MOD11C1, MYD11C1, MYD11A1 MOD13C1, MCD15A3). Among them, MOD11C1 and MYD11C1 provide daily land surface temperature at 0.05 spatial resolu-5 tion, while MYD11A1 provides daily surface temperature at 1 km resolution. MOD13C1 contains 16 day composite of Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI). MCD15A3 provides the combined (Terra and Aqua) MODIS Leaf Area Index (LAI) and Fraction of Photosynthetically Active Radiation absorbed by vegetation (FPAR) products every 4 days at 1 km resolution. The above prod- 10 ucts have all been validated against a wide range of in situ observations, and applied in many scientific studies (Coll et al., 2009;Fensholt et al., 2004;Tian et al., 2002).

MSG SEVIRI data
The Spinning Enhanced Visible and Infrared Imager (SEVIRI) is the main instrument on board Meteosat Second Generation (MSG) geostationary satellites (Schmetz et al., 15 2002). The SEVIRI has 12 separate channels from visible to thermal infrared and can provide observations at very high temporal resolution (every 15 min), which makes it possible to resolve the diurnal cycle of environmental variables such as land surface temperature (Peres and DaCamara, 2004). So far, geostationary satellite data has not been used for soil moisture downscaling. Compared to the polar orbit satellites, the 20 geostationary satellites normally provide measurements with relatively low spatial resolution. As for SEVIRI, its spatial resolution is approximately 4.8 km with spatial sampling of 3 km for nadir view. But the major advantage of the SEVIRI over MODIS is the high temporal frequency (96 times per day), which can highly increase the possibility of obtaining cloud-free measurements. Furthermore, the downscaling approach used in this 25 study can further benefit from the increased observation frequency through obtaining thermal inertial information from surface temperature diurnal cycle. It should be noted here that the satellite-based products used in this study have different data formats, spatial resolutions and projections. Therefore, a preprocessing is required to make them consistent in space. The surface temperature and vegeta-5 tion indexes products from MODIS and SEVIRI are all resampled to a regular latitude/longitude grid with 0.05 • spacing.

Methodology
The methodology used in this study includes a soil moisture downscaling scheme and evaluation strategies. The details are described in the following paragraphs.

Downscaling scheme
The soil moisture downscaling scheme used in this study was proposed by Peng et al. (2015). It uses VTCI as the only scaling factor to improve the soil moisture from coarse to high spatial resolution. The theoretical basis of this approach is that the VTCI can represent the status of soil moisture, and has been widely used for estimating soil 15 moisture and monitoring drought condition (Petropoulos et al., 2009). The estimation of VTCI is based on the triangular or trapezoidal feature space that is constructed by land surface temperature and vegetation index (Fig. 2) over the study area. It is calculated by rescaling the surface temperature of each pixel between two extreme surface temperature values for each vegetation index interval: where T s is the observed surface temperature for a given pixel, T max and T min are the corresponding maximum and minimum surface temperatures which have the same 8514 Introduction

Tables Figures
Back Close

Full Screen / Esc
Printer-friendly Version

Interactive Discussion
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | vegetation index value as the given pixel. The assumption behind is that the variation of T s between the two extreme T s values reflects the changes of evapotranspiration and soil moisture (Peng and Loew, 2014). The maximum and minimum temperature values form the dry and wet edges in the triangular/trapezoidal feature space. The dry edge reflects the status of limited soil moisture and minimum evapotranspiration, while 5 the wet edge reflects the condition of unlimited soil moisture and maximum evapotranspiration. To accurately estimate dry and wet edges, different approaches have been proposed such as Tang et al. (2010) and Long et al. (2012). Some of these methods are physically based, but increase the operational difficulty due to the requirements of ground-based measurements such as air temperature and wind speed. To keep the performing the linear fit, the maximum surface temperature that are to the left (with lower vegetation index) and the maximum surface temperature less than mean minimum surface temperature are removed, to filter out the spurious dry points and outliers. The wet edge is the mean of the minimum surface temperature within the last five vegetation index intervals.

20
In order to downscale the coarse resolution CCI soil moisture, the spatially average VTCI (VTCI) is then estimated for each CCI soil moisture grid box at 0.25 • (CMG) as: where n is the number of 0.05 • grids in the SM 0.25 • grid. Based on the above results, the following equation is used to downscale the soil moisture: where SM is the downscaled CCI soil moisture at 0.05 • (CMG), SM is the actual CCI soil moisture at 0.25 • (CMG), VTCI stands for the scaling factor at 0.05 • (CMG). Com-5 pared to other downscaling methods, this downscaling approach is very simple and requires only limited input data. Since VTCI is the key variable in the downscaling scheme, it determines the accuracy of the downscaled soil moisture field. The performance of VTCI mainly depends on the accuracy of surface temperature and type of vegetation index. Furthermore, 10 clouds, terrain and land cover heterogeneity might be also sources of the errors for the estimation of VTCI (de Tomás et al., 2014). Therefore, to optimize the performance of VTCI as the proxy of soil moisture, a sensitivity analysis of VTCI to five influential factors (surface temperature, vegetation index, cloud, topography and land cover heterogeneity) was conducted before downscaling soil moisture. The VTCI was firstly 15 calculated with different settings (see Table 3). Then the estimated VTCI were compared with ground-based soil moisture measurements from the REMEDHUS network to investigate the influence of the five factors on the performance of VTCI. The details of these factors are listed below, and the combinations of these factors for estimating VTCI are shown in Table 3. 20 1. Surface temperature: the instantaneous surface temperature observed at the time of satellite overpass is typically used for constructing a triangular/trapezoidal feature space. To avoid the influence of uncertainty of instantaneous surface temperature and make use of thermal inertial information, the surface temperature difference has been used by many studies such as Stisen et al. (2008)   4. Topography: the estimation of VTCI requires a relatively flat area, because the variation of surface temperature in the triangular/trapezoidal feature space is assumed to be caused by evaporative cooling effect rather than the variation of elevation. To examine the effects of topography on VTCI, the performances of non-masked terrain and masked terrain are both tested. In this study the masked terrain is based on removing the areas with 300 m higher or lower elevation than average REMEDHUS elevation.

5.
Land cover heterogeneity: in order to get the triangular/trapezoidal feature space, the study domain should cover a wide range of soil moisture and vegetation cover conditions. On the other hand, the study area also needs to be homogeneous in terms of vegetation type and surface roughness (Moran et al., 1994). The sensitivity of VCTI to land cover heterogeneity is evaluated at two conditions: (1) full 5 land cover, (2) only cropland.
Finally, the optimized VTCI is calculated based on the sensitivity analysis results. The CCI soil moisture is then downscaled with the optimized VTCI.

Evaluation strategies
Two metrics are used to evaluate the performance of the downscaled soil moisture. 10 The first is direct comparison between satellite-based products including CCI SM and downscaled CCI SM, and measured soil moisture at each REMEDHUS station. In addition, the cross comparisons between the present results and reported results from published researches are summarized as well. In order to investigate the influence of land use on the downscaled soil moisture, the comparisons between satellite-based re- 15 sults and in situ measurements are performed per land use of the stations. These land use are rainfed, vineyard and forest-pasture. Furthermore, a seasonal analysis is also carried out to examine its influence on downscaled soil moisture. The study period is separated into four seasons to represent different dry/wet conditions and vegetation growth conditions. The four seasons are September/October/November (autumn),

20
December/January/February (winter), March/April/May (march), and June/July/August (summer) respectively. The widely used statistical metrics including correlation coefficient (R), mean bias error (BIAS), root mean square difference (RMSD), and unbiased RMSD (ubRMSD) are used in this study to quantify the differences between satellitebased products and in situ measurements.  Figure 3 shows the box plots of R for the comparisons between in situ soil moisture measurements and VTCI estimated from different settings (Table 3) for all stations. For 5 the effects of surface temperature, the use of surface temperature difference generally shows better results than the use of instantaneous surface temperature. However, for SEVIRI, the use of temperature difference between 8:00 and 12:00 a.m. does not improve VTCI compared to the use of instantaneous surface temperature at 10:30 a.m. and 13:30 p.m. For Terra MODIS, the daytime nighttime temperature difference method 10 performs slightly worse than the daytime method. It might be caused by the use of high quality surface temperature products in this study. As stated before, one advantage of temperature difference method is avoiding the uncertainty of instantaneous surface temperature. The similar performance between temperature difference method and instantaneous temperature method can be expected if the surface temperature product 15 has high accuracy. Nevertheless, the improvements of VTCI are clearly observed when using maximum and minimum temperature difference for SEVIRI, as well as daytime nighttime temperature difference for Aqua MODIS. These results indicate the effectiveness of integrating the information of thermal inertial. From definition, the thermal inertial requires the maximum diurnal temperature difference, which is fully met by the 20 SEVIRI. For Aqua MODIS, the observations at 13:30 a.m. and 13:30 p.m. are also close to the local maximum and minimum temperature. But in reality, the time of local maximum and minimum temperature change with seasons and locations. Therefore, the use of SEVIRI can give better performance than MODIS platforms due to the better integration of thermal inertial information. It should be also noted that Aqua has slightly 25 better performance than Terra in terms of mean R. The change of VTCI when using different type of vegetation index is also shown in Fig. 3. The LAI gives the best performance in terms of mean R for both SEVIRI and MODIS platforms, and then followed by EVI and FPAR. The NDVI gives relatively worse performance. It is because NDVI is only an indicator of surface greenness due to its saturation in dense vegetation and sensitivity to effect of soil background. The other 5 indexes do not suffer these problems and can better represent the reality of vegetation density, which leads to better performance of VTCI. To our knowledge, many studies have used one of these vegetation indexes to form the triangular/trapezoidal feature space, but none of them compare and quantify the difference of VTCI estimated from different types of vegetation indexes. The results suggest that LAI is the best proxy for 10 vegetation cover among different vegetation indexes in the applications based on the triangular/trapezoidal feature space.

HESSD
Regarding the influence of clouds, Fig. 3 shows that the 85 % cloud mask gives worse performance for VTCI than 75 % cloud mask. In theory, the increase of the cloud mask threshold should lead to the improvement of the accuracy of VTCI. The opposite 15 result obtained here is due to the sharply decreased sample days for 85 % cloud mask. It should be noted that the totally clear sky in the study domain is rare in real conditions. The higher cloud mask threshold normally results in less sample days. To keep the balance between avoiding the influence of clouds and having more sample days, we use 75 % cloud mask in this study.

20
Contrary to our expectations, the masked terrain performs quite similar to nonmasked terrain method. It is because that the masked out pixels are normally located within the triangular/trapezoidal feature space (see green color points in Fig. 5f and g), which means the dry and wet edges keep almost the same for both methods. Therefore, the terrain has no strong impacts on VTCI in our study area. But for other study 25 areas, the terrain effects still need to be investigated before the estimation of VTCI.
The use of full land cover types gives better performance of VTCI than the use of croplands. It suggests that although the use of croplands keeps the surface cover more homogeneous in terms of vegetation properties and surface roughness, the range of surface moisture conditions that required for the estimation of VTCI decreases meanwhile. In arid or semiarid study areas such as our study area, only one vegetation type cannot represent a wide range of soil moisture that required by the estimation of VCTI. For these kinds of study areas, the requirements of homogeneous surface cover and wide range of soil moisture conditions cannot be met at the same time. As shown in 5 our results, for the estimation of VTCI, having a wide range of soil moisture is more important than keeping surface cover homogeneous in such areas. Based on the above results, it can be seen that these factors have strong impacts on performance of VTCI. The optimal configurations in this study for the estimation of VTCI are using Aqua MODIS daytime nighttime temperature difference, SEVIRI maximum 10 and minimum temperature difference, LAI, 75 % cloud mask, non-masked terrain and full land cover types.

VTCI as a proxy of soil moisture
To further investigate the performance of VTCI as the proxy of soil moisture, the temporal evolution of LAI, surface temperature, VTCI and in situ soil moisture over REMED- 15 HUS are presented in Fig. 4. Meanwhile, the R between each parameter and soil moisture measurement is also shown. The LAI has similar seasonal trend as in situ measured soil moisture, but with very low R of 0.04 between them. Compared to LAI, a significantly negative correlation with R of −0.24/−0.56 for MODIS/SEVIRI between surface temperature and soil moisture are obtained. The results suggest that both LAI 20 and surface temperature can represent the status of soil moisture, and the surface temperature is more sensitive to soil moisture than the LAI. As expected, the VTCI, combining the information from both LAI and surface temperature, agrees well with soil moisture with R of 0.37/0.52 for MODIS/SEVIRI, which further demonstrating the effectiveness of VTCI as a proxy of soil moisture.

Spatial patterns of the soil moisture estimates
On the basis of VTCI, the CCI SM is downscaled to high spatial resolution using the proposed method during the study time period. The spatial distributions of original CCI SM and downscaled soil moisture on 22 May 2010 are shown in Fig. 5. It can be clearly seen that the downscaled soil moisture (Fig. 5b and c) have quite similar spatial 5 patterns as the original CCI SM. High soil moisture is typically presented in northwest and southwest, while low soil moisture appears in northeast and southeast. Meanwhile, the spatial details of the soil moisture are highly improved by the downscaling scheme. The downscaled soil moisture map (Fig. 5d) generated from MODIS also exhibits very similar patterns as that (Fig. 5e) from SEVIRI. It is due to the similar VTCI patterns calculated from MOIDS and SEVIRI. The similar shape of the triangular/trapezoidal feature space constructed from MODIS and SEVIRI can also be seen from Fig. 5f and g. These results suggest that the proposed downscaling scheme can capture the spatial pattern of original CCI soil moisture, and similar performance of downscaled soil moisture maps can be obtained from both MODIS and SEVIRI. The following section 15 will further investigate the accuracy of the downscaled soil moisture and quantify the difference between estimates from MODIS and SEVIRI.

Validation of the soil moisture estimates against in situ soil moisture measurements
The validation results between original CCI SM and in situ soil moisture measure-  stations like L3V and K4V. These results suggest that the downscaled soil moisture can maintain the accuracy of original CCI SM but cannot highly improve its accuracy. Similar results have been reported by many soil moisture downscaling studies (Choi and Hur, 2012;Sánchez-Ruiz et al., 2014;Zhao and Li, 2013), which suggest that the accuracy level of downscaled soil moisture highly depend on the original soil moisture. Besides, as expected, slightly better performance of SEVIRI over MODIS can be observed due to the better performance of VTCI from MSG, which demonstrates the potential of MSG for downscaling soil moisture. The temporal variations of downscaled soil moisture for individual station are also investigated here. The stations K13 and M13 are selected due to their representatives 15 of wet and dry soil moisture conditions (Saìnchez et al., 2012). Besides, the location of K13 is close to one weather station (Fig. 1), which gives us the chance to investigate the connection between soil moisture and rainfall. Figure 7 displays the time series of the in situ soil moisture, CCI soil moisture, downscaled CCI soil moisture, as well as rainfall for stations K13 and M13 respectively. For dry station K13, the CCI soil moisture 20 and downscaled soil moisture both agree well with in situ soil moisture. Regarding the wet station M13, the in situ soil moisture responds well to the rainfall, with high soil moisture occurring during the rainfall period in spring. But the CCI soil moisture and downscaled soil moisture seem to be insensitive to the rainfall, presenting relatively low value compared to in situ soil moisture during the rainfall period. The results here are 25 similar to that reported by (Sánchez-Ruiz et al., 2014). They found that the downscaled SMOS soil moisture has limited response to rain events.
In addition to the validation at each individual station, the performance of different soil moisture results averaged at REMEDHUS network scale are also analyzed and HESSD 12,2015 Evaluation of soil moisture downscaling using a simple thermal based proxy  Fig. 8. Similar to the above results, both original CCI SM and downscaled soil moisture agree well with the averaged in situ soil moisture over network in terms of R, BISA, RMSD and ubRMSD. But systematic overestimation of soil moisture from all of them can also be observed. It implies that the CCI SM has the problem of overestimating soil moisture, which needs to be further investigated. Overall, the 5 above results suggest that the downscaled soil moisture can preserve the accuracy of coarse CCI SM, and meanwhile present more detailed spatial details. Since the proposed downscaling method highly depends on the original CCI soil moisture and VTCI, the accuracy of the downscaled soil moisture is expected to be highly improved if the VTCI can perfectly represent the soil moisture.
Furthermore, the above results are also compared with other published soil moisture downscaling studies. These studies apply different downscaling methods to downscale soil moisture product from SMOS and AMSR-E, using either MODIS or SEVIRI data as inputs. The results are all validated against the observations from REMEDHUS network, which makes them ideal for inter-comparison with our results. Table 4 lists the 15 statistics of the comparison between downscaled and measured soil moisture from different studies. It can be seen that the R, BIAS, RMSD and ubRMSD of these published studies range respectively from 0.467 to 0.73, −0.026 to 0.108, 0.049 to 0.109, and 0.040 to 0.042 m 3 m −3 . The corresponding statistics of the current study are within theses ranges, which imply that the results of current study are reasonable and sat-20 isfactory. Considering the simplicity, limited inputs, and acceptable accuracy, we conclude that the proposed downscaling method is feasible and effective for improving soil moisture from coarse to high spatial resolution.

Seasonal and land use analyses of the soil moisture estimates
To investigate the performance of the downscaling method over different climatic and 25 vegetation growth conditions, seasonal analysis has been performed based on the comparisons between network-averaged soil moisture estimates and in situ soil moisture observations for different seasons. Figure 9 shows the statistical results of the 8524 response to rainfall, while the study area has frequent rainfall during spring. Compared to CCI soil moisture, the downscaled soil moisture from SERVIRI has slightly better performance in terms of R and BIAS. Similar to the previous results, the SERVIRI generally has slightly better performance than MODIS in terms of R, BIAS, RMSD and ubRMSD values. 10 The influence of land us on the downscaling scheme is also investigated and the results are summarized in Fig. 10. The stations are divided into three land use groups: vineyard (7), rainfed (9) and forest-pasture (2). Figure 10 shows that different land use categories have similar performance, and slightly better performance is obtained by forest-pasture in terms of R. These results suggest that both CCI SM and downscaled soil moisture are not sensitive to the generally used land use. Furthermore, the results from MODIS and SERVIRI have similar performance, which further implies that the proposed method is independent on platforms. It has the potential to be used on different platforms and other regions.

Evaluartion of downscaled soil moistrue at different spatial resolutions 20
Since MODIS has the advantage of providing measurements at high spatial resolution of 1 km, it gives us the opportunity to evaluate the downscaled soil moisture at different spatial resolutions. Figure 11 displays the spatial patterns of downscaled soil moisture at 1 km and 0.05 • (CMG) on 22 May 2010. It can be seen that these two maps have quite similar patterns, implying that the downscaling scheme can be applied at differ-25 ent spatial scales. To further evaluate the performance of downscaled soil moisture at different spatial resolutions, Fig. 12 shows the bar plots of the comparisons between downscaled soil moisture from MODIS different spatial resolution MODIS and mea-8525 Introduction

Tables Figures
Back Close

Full Screen / Esc
Printer-friendly Version

Interactive Discussion
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | sured soil moisture at each station. In general, the downscaled soil moisture at 1 km has similar accuracy level as that at 0.05 • , which suggests that the accuracy of CCI soil moisture can be preserved at higher spatial resolutions. Besides, the 1 km soil moisture has slightly better performance than 0.05 • soil moisture in terms of mean R, RMSD, ubRMSD values. It further suggests that finer spatial resolution datasets can 5 improve the accuracy of downscaled soil moisture, which demonstrates the assumption proposed by Sánchez-Ruiz et al. (2014). In summary, the above results indicate that the downscaling scheme can be applied at different spatial resolutions. Taking advantage of high spatial resolution of MODIS datasets, combined use of MODIS and MSG datasets has the potential of providing downscaled soil moisture at high spatial 10 and temporal resolutions.

Conclusions
In this study, a newly developed soil moisture downscaling method was applied to the ESA CCI soil moisture product and validated against the REMEDHUS soil moisture observation network in Spain. In general, agreement between the CCI soil moisture 15 and the in situ soil moisture was observed with similar accuracy level to the published validation studies. But systematic overestimation of soil moisture was also observed for CCI SM from the network-averaged analysis. Before applying the downscaling scheme, the sensitivity analyses of the downscaling factor VTCI were conducted. The surface temperature difference method performs better than instantaneous surface tempera-20 ture method due to the integrated information of thermal inertial. Besides, the VTCI performance also depends on the type of vegetation index. The LAI performs best for estimation of VTCI compared to NDVI, EVI and FPAR. After downscaling the CCI soil moisture from coarse to high spatial resolution, the soil moisture map can replicate the CCI soil moisture spatial patterns and show more spatial details. downscaling studies suggest that the accuracy level of the proposed method is comparable. Compared with those methods, the advantages of the proposed method are simple, less inputs and comparable accuracy level. In addition, the downscaled soil moisture from MSG SEVIRI performs better than that from MODIS, which is due to the better performance of corresponding VTCI from 5 SEVIRI. It indicates the great potential of applying SEVIRI to downscale soil moisture. To take full advantages of the high temporal resolution of SEVIRI and high spatial resolution of MODIS, combined use of data from both platforms should be considered in soil moisture downscaling applications in the future.
In summary, the present study, together with the work by Peng et al. (2015) demon-10 strated the feasibility of downscaling soil moisture with the proposed method. The notable advantage of this approach is simplicity in terms of inputs requirement and implementation. Furthermore, the proposed method is independent on satellite platforms, implying that the downscaled soil moisture can be obtained at either very high spatial resolution (500 m for MODIS) or very high temporal resolution (every 15 min for 15 SEVIRI). It will facilitate regional hydrological related studies that require soil moisture information at different spatial and temporal scales.
In addition, the authors would like to thank the International Soil Moisture Network (ISMN) for making the in situ measurements over REMEDHUS public available. The authors also would like to thank José Martínez-Fernández from Universidad de Salamanca for providing rainfall measurements from weather stations installed in the REMEDHUS observation network.

Tables Figures
Back Close

Full Screen / Esc
Printer-friendly Version