0 Introduction

Surface soil moisture exists within the interface of the atmosphere, pedosphere, and biosphere.It plays a major role in the distribution of sensible heat and latent heat in solar net radiation, as well as in the runoff and infiltration process during precipitation. It provides important moisture source to cloud for precipitation by evaporation^[1]. It also plays the determined role in dynamic variation of deep soil moisture^[2]. Therefore, it is a critical initial parameter for simulating land surface processes such as climate and ecology.

Because the soil moisture is strongly variable in space and time, traditional in situ station monitoring methods are ineffective in assessing soil moisture on regional spatial or wide temporal scales; while microwave remote sensing provides an effective alternative for long-term real-time dynamic monitoring of soil moisture^[3-4]. As the first passive microwave sensor to provide a global soil moisture product, the AMSR-E has been being widely used. For example, Draper et al.^[5] stated that soil moisture datasets by AMSR-E accurately reported the soil moisture in Australia. Zahid and Rasul^[6] used AMSR-E data to analyze the spatial and temporal variation of soil moisture during crop planting season of Pakistan from 2003 to 2010. Chen et al.^[7] verified the reliability of the AMSR-E soil moisture product in the Xilinhot grassland plots in China. Xi et al.^[8]compared the accuracies of the three AMSR-E soil moisture products (JAXA, NASA, and VUA) in the Qinghai-Tibet Plateau (QTP), and reported that NASA and VUA presented higher retrieval accuracy. However, the official dynamic range of soil moisture was small, and therefore, which could not reflect any inter-annual trends.

Several studies have sought to improve the AMSR-E original retrieval algorithm by introducing the Q_p surface radiation model^[9], and to validate this improved algorithm and the retrieval accuracy using in situ measured data^[10-11]. In addition, regression analysis relating in situ data to the remote sensing products reduces the error in soil moisture estimation with the AMSR-E ^{[11, 12]}. For instance, Feng et al.^[13] used monthly regression analysis and the results on the soil moisture variations over the Poyang Lake basin in China were improved.

In order to determine the most effective method, we intend to compare and validate the various AMSR-E soil moisture products in this study, including those retrieved by the dual-channel retrieval algorithm with the Q_p model (SM_D), those using the linear regression-corrected method (SM_L), and the AMSR-E official soil moisture (SM_O). We will then apply these results to the analysis of the spatial and temporal variation in soil moisture over the SRYR in China. The SRYR, located in the Qinghai-Tibet Plateau (QTP), is also the world's most important high-altitude biodiversity nature reserve^[14]. In recent decades, particularly as a result of climate change and human activity in the area, the SRYR has experienced environmental deterioration, including the loss of runoff flow in the main river, the shrinking of associated lakes, soil erosion, continued degradation of area wetlands, deterioration of adjacent grassland ecosystems, and increasing desertification. These problems have had a serious impact on the local ecology, economics such as livestock production, and sustainability of water resources within the Yellow River basin^[15-17].

Soil moisture is crucial to many ecological processes in alpine grasslands, including the ecological carrying capacity, grassland resilience, and grassland recovery and reconstruction from degradation^[18]. Soil moisture can cause variations in atmospheric heat content, which would impact the shift in seasons in the SRYR^[19]. Generally, few studies have examined the long-term or large-scale soil moisture dynamics as they relate to climate change. This research uses the AMSR-E soil moisture products, along with in situ data, to examine the spatial and temporal variations of soil moisture in the SRYR, to assess the contributing factors to these variations, and to understand the response of soil moisture to climate change as well as regional ecological and hydrological processes. Finally, this research assesses water resources management and ecological restoration options for the study area.

1 Materials and methods 1.1 Study area

The SRYR is a catchment with an area of approximately 145 300 km² above the Longyangxia Reservoir in the mainstream of the Yellow River, and its geographic range is between 32° 10′ and 36° 59′ N in latitude and between 95° 54′ and 103° 24′ E in longitude (Fig. 1). The climate in SRYR is a typical continental plateau, with an average annual air temperature of-3 to-4.1 ℃ and an annual average precipitation of 300-700 mm. The landform of the Yellow River is mainly three basic types of the mountains with an average altitude of > 4 000 m, hill terraces and plain, with undulating plateau planation surface as the main form. The underlying surface primarily consists of seasonal permafrost, alpine meadows, alpine swamps, alpine lakes, and wetlands^[20]. Based on the division of the Qinghai-Tibetan Plateau natural zone^[21], the main regions of the SRYR are as followings: the wide valley basin of the YRSR (Ⅰ), Zoige hummocky plateau (Ⅱ), Golog Yushu plateau gully (Ⅲ), Huangnan mountains (Ⅳ), and the eastern margin of the Qaidam mountains (Ⅴ).

1.2 AMSR-E data of soil moisture

The AMSR-E L3 daily surface moisture data during June 2002 to October 2010 from the National Snow and Ice Data Center (NSIDC) was applied for this study, AMSR-E was on the Aqua satellite launched by NASA in 2002. It had 6 observation channels of 6.9, 10.7, 18.7, 23.8, 36.5, and 89.0 GHz, each with dual vertical and horizontal polarized radiation measurements, totally 12 channels. The data were available twice daily from an ascending track (overpasses at 13:30 local time) and a descending track (overpasses at 1:30 local time) with EASE-Grid as projection method and a spatial resolution of 25 km. Because the AMSR-E data in each day cannot cover a full day in the middle and low latitudes, we spliced the data from two adjacent days into one using the stitching algorithm.

1.3 Measured soil moisture, meteorological data and NDVI

The soil moisture data from the Maqu Soil Moisture Monitoring Network (MSMMN) located in the SRYR was utilized to calibrate and validate the AMSR-E soil moisture product downloaded from the International Soil Moisture Network. The soil moisture was measured at different depths (5 cm, 10 cm, 30 cm, 50 cm, and 80 cm below the surface) at 15-min intervals by the EC-TM ECH₂O probe (Decagon Devices, Inc., USA).The EC-TM ECH₂O is a capacitance sensor that measures the dielectric permittivity of the soil surrounding the probe's pins, and the root mean square error (RMSE) was 0.02-0.06 m³/m^3[22]. Since microwave can only reach a few centimeters deep in the ground, the soil moisture at the surface of the 5 cm observed in this paper will be used for analysis and verification. The MSMMN is located at Zoige hummocky plateau of the Southeast SRYR, near to the first big bend of the Yellow River, Maqu County of Gansu province; the underlying surface is alpine meadow. The MSMMN consisted of 20 sites (as shown in Fig. 1 and Tab. 1). The data from 8 sites of CST-01, CST-04, NST-02, NST-04, NST-07, NST-10, NST-12, and NST-13 were used for the calibration of AMSR-E soil moisture product, while the data from other 12 sites for the verification. In order to obtain the actual surface soil moisture during the passage of the satellite, the average measured soil moisture adjacent to the satellite transit time was chosen to represent the surface soil moisture of AMSR-E transit time.

Mean monthly air temperature and precipitation data were obtained between 2002 and 2011 from 11 National Meteorological Information Center (http://data.cma.cn) stations located across the SRYR (as shown in Fig. 1). Using ARCGIS10.2, we conducted a Kriging interpolation of the air temperature and precipitation point data, and then generated raster image datasets with the same spatial resolution and soil moisture by pre-processing such as projecting, clipping, and re-sampling.

MOD13A3 is monthly vegetation index developed by the land group of NASA MODIS via a common algorithm, and can be downloaded from Geospatial Data Cloud (http://globalchange.nsdc.cn), and the data were pre-processed by radiation, geometrical and atmosphere calibration. The data for this study was the maximum monthly NDVI through the Maximum Value Composite (MVC) method with spatial resolution of 1 km. The re-sampled data calibrated by projection matched the data of soil moisture.

1.4 Dual-channel retrieval algorithm with Q_p model

In the retrieval of surface soil moisture from passive remote sensing data, how to remove the influence of the surface roughness is an important issue. Shi et al^[9] solved this problem through the development of the Q_p model. The Q_p model is a surface radiation model aimed at AMSR-E sensor parameters based on the Advanced Integral Model (AIEM), applied to high frequency and wide surface roughness. It can be expressed as:

${e_{\rm{p}}} = {Q_{\rm{p}}}{t_{\rm{q}}} + \left( {1-{Q_{\rm{p}}}} \right){t_{\rm{p}}}$

(1)

where e_p is the rough surface emissivity (equivalent to e_v and e_h in formula 2 shown below), t_q is the polarized fresnel transmittance, t_p is the fresnel transmittance, and Q_p is the roughness parameter. The subscript p represents polarization (vertical or horizontal polarization).

In this study, the SM_D dataset is processed by a dual-channel retrieval algorithm with Q_p model developed by Shi et al.^[9]. This algorithm is based on a single-channel retrieval algorithm developed by Jackson et al.^[23], and calculates the land surface temperature on the basis of a 36.5-GHz vertical polarized brightness temperature^[24]. The algorithm eliminates the influence of vegetation on the microwave signal by determining the vegetation's optical thickness using the empirical relationship between the vegetation water content and the NDVI. The dual-channel retrieval algorithm with Q_p model is then introduced into the model in order to eliminate the influence of land surface roughness. The resulting soil moisture data is then taken as the inverse of the 10.65-GHz brightness temperature. The algorithm^[11] can be expressed as follows:

$S{M_{\rm{D}}} = 5.50 + 1.15\left( {2.32{e_{\rm{v}}} + {e_{\rm{h}}}} \right)-5.13\sqrt {2.32{e_{\rm{v}}} + {e_{\rm{h}}}} $

(2)

where the SM_D is the retrieved soil moisture data and e_v and e_h are the vertical and horizontal polarized rough surface emissivity, respectively, calculated as the ratio between the brightness temperature and land surface temperature.

1.5 Monthly regression analysis

Because the AMSR-E soil moisture retrieval algorithm is strongly influenced by seasonal changes in vegetation cover, applying a monthly regression analysis to the AMSR-E soil moisture product increased the accuracy of the soil moisture measurements.In the monthly regression analysis, the in situ data was used to calibrate the AMSR-E soil moisture^[13]. The regression can be expressed as:

$S{M_{\rm{L}}} = aS{M_{{\rm{AMSRE}}}} + b$

(3)

where SM_L is the calibrated final product, and SM_AMSRE is the original AMSR-E soil moisture product, while a and b are the regression coefficients.

1.6 Data analysis

The trend of the soil moisture during 2003-2010 served as the changing rate representing drier or wetter conditions, annually. We used the trend line method to analyze the change in soil moisture in different areas and months during 2003-2010, with the slope of the trend line calculated as:

${\theta _{{\rm{Slope}}}} = \frac{{n\sum\limits_{j = 1}^n {\left( {j{x_j}} \right)}-\left( {\sum\limits_{j = 1}^n j } \right)\left( {\sum\limits_{j = 1}^n {{x_j}} } \right)}}{{n\left( {\sum\limits_{j = 1}^n {{j^2}} } \right)-{{\left( {\sum\limits_{j = 1}^n j } \right)}^2}}}$

(4)

where n is the number of years (for this study, n = 8, year 2003-2010), x_j is the soil moisture value in jth year, and θ is the slope of the trend line (θ > 0 indicates that the change in the soil moisture in n years is increased).

In this study, the validation accuracy of three soil moisture products has been proposed based on observed data. The coefficient of determination (R²), between in situ data and the soil moisture product, and the RMSE were computed as criteria for goodness of fit. The RMSE is defined as:

${\rm{RMSE = }}{\left( {\frac{{\sum\limits_{i = 1} {{{\left( {S{M_{{\rm{in}}\;{\rm{situ}}}}-S{M_{\rm{v}}}} \right)}^2}} }}{n}} \right)^{1/2}}$

(5)

where SM_{in situ} and SM_v are the observed and validated values for ith pair, and n is the total number of paired values. Smaller RMSE values corresponded to smaller differences between the validated values and the observed values.

2 Results 2.1 Calibration and verification of AMSR-E soil moisture data

First, according to the dual-channel retrieval algorithm with Q_p model, we calibrated the AMSR-E soil moisture using in situ soil moisture from eight sites in the MSMMN, and then validated them with data from the other 12 sites. Tab. 2 shows that the accuracy of SM_D is higher than SM_O, and the RMSE is reduced from 0.041 to 0.036. At the same time, using the NST-14 site in 2009 as an example (see Fig. 2), the range of SM_O is small (0.070-0.180 cm³/cm³), and SM_D was significantly higher than the in situ soil moisture in general situation, especially so in the months with little moisture. SM_D is generally lower than the in situ data, and in periods with less precipitation and vegetation, SM_D is similar to the in situ data. However, in periods with more precipitation and vegetation, SM_D is much lower than the in situ data. Tab. 2 shows that the AMSR-E soil moisture retrieval algorithm is not strongly influenced by precipitation, but rather seasonal changes in vegetation cover^[3]. Applying monthly regression analysis to the AMSR-E soil moisture product could mitigate the effect of the change in vegetation cover for increasing the accuracy of the soil moisture measurements.

Next, we calculated the monthly calibration coefficient using equation (2) based on SM_D bydual-channel retrieval algorithm with Q_p model as well as in situ data from eight sites, and obtained the soil moisture product (SM_L) using the linear regression-corrected method. Then, the accuracy of SM_L was validated by data from the 12 sites. Tab. 2 shows that the accuracy of the SM_L is higher than SM_D and SM_O. The range and trend curve of SM_L most closely approximate those of in situ data (see Fig. 2). Consequently, we calibrated the AMSR-E soil moisture through the dual-channel retrieval algorithm with Q_p model and monthly regression analysis, and analyzed the spatial and temporal variations in soil moisture over the SRYR from 2003 to 2010.

2.2 Spatial and temporal variations in soil moisture

Average annual land surface soil moisture is 0.140-0.380 cm³/cm³ as shown in Fig. 3, the overall spatial distribution presented as high in the southeastern part and low in the northwestern part of the study area, i.e., gradually rising from northwestern to southeastern part. It is similar to the soil moisture calculated for the Qinghai-Tibet Plateau^[11], but smaller than the range of 0.2-0.5 cm³/cm³ in the SRYR from August to October 2009 by the Advanced Synthesis Aperture Radar (ASAR)^[25]. The average soil moisture in the Zoige hummocky plateau of the Southeast SRYR is 0.332 cm³/cm³, and this is the highest in the SRYR, and the area near the first bend of the Yellow River is of the highest soil moisture, and it is a key water conservation area in the SRYR. Correspondingly, Chen et al.^[26]found that the soil moisture in the Zoige wetland was approximately 0.3 cm³/cm³ from their soil moisture simulation experiment. The average annual soil moisture in the Golog Yushu plateau gully of the South SRYR is 0.275 cm³/cm³ with the trend of high in the southeast and low in the northwest of the area. The soil moistures in the Huangnan mountains of the North SRYR, and the wide valley basin of the Northwest YRSR are low at averagely 0.215 cm³/cm³ and 0.241 cm³/cm³, respectively. The soil moisture in the eastern margin of the Qaidam Mountains in the most northern part of SRYR is the lowest at 0.213 cm³/cm³.

The soil moisture tends to decrease from 2003 to 2010 in the SRYR at a rate of 0.012 cm³/cm³ yearly (see Fig. 4). There is also a slight trend of drying in the northern Tibetan over a period of nearly ten years^[27]. Furthermore, the soil moisture in different natural zonation is inconsistent. In the Zoige hummocky plateau, Golog Yushu plateau gully and Huangnan mountains, the soil moisture gradually decreases at a rate of 0.067 cm³/cm³, 0.011 cm³/cm³, and 0.012 cm³/cm³, respectively. In the wide valley basin of the YRSR and the eastern margin of the Qaidam mountains, the soil moisture increases at a rate of 0.026 cm³/cm³ and 0.019 cm³/cm³, respectively.

Fig. 5 shows the spatial variations in the soil moisture averaged across the study period (2003-2010). Regarding to varied seasons and areas, the soil moisture in the spring (March, April, and May) decreases in most areas, especially in the Zoige hummocky plateau and in the Huangnan mountains; while the spring soil moisture in the wide valley basin of the YRSR and in the Golog Yushu plateau gully shows an increasing trend. In the summer (June, July, and August), the soil moisture increases in the eastern margin of the Qaidam Mountains and in the western Eling Lake in the wide valley basin of the YRSR, and decreases in the southern area of both Golog Yushu plateau gully and Huangnan mountains. In the autumn (September, October, and November), there are significant fluctuations in the soil moisture in the southern Zoige hummocky plateau and Golog Yushu plateau gully. In the winter (January, February, and December), the SRYR experiences freezing soils, as the mean air temperature drops below zero at times, resulting in minimal fluctuations in the soil moisture. On an individual monthly basis, the averaged soil moisture gradually increases in June, July, and September (when the moisture shows a significant increase of 0.036 cm³/cm³. In all other months, the averaged soil moisture decreases, with the greatest rate occurring in May (0.009 cm³/cm³ and November 0.008 cm³/cm³.

2.3 Analysis of factors influencing soil moisture

Tab. 3 shows a significant positive correlation between precipitation, NDVI, and soil moisture in the SRYR, while air temperature is not correlated with the annual averaged soil moisture. Using monthly data, there is lower correlation between the precipitation and soil moisture in the January and December than the other months, likely due to hindered infiltration when the ground is frozen. From April to September inclusive, air temperature and soil moisture show a negative correlation, particularly in May and June when the soil moisture is most strongly subjected to evaporative influences from high ambient temperatures. From January to March and from October to December, air temperature and soil moisture are positively correlated, likely because in the freeze-thaw cycles when increased air temperatures facilitate the thaw of accumulated snow and frozen water, therefore increasing the soil moisture. The NDVI is also significantly correlated with the soil moisture, and it is consistent with previous studies^{[10, 28]}. Vegetation affects the soil moisture by intercepting rainfall as well as regulating evaporation and infiltration. The correlations in May to July are generally lower than the other months, likely as the effect of air temperature on the soil moisture during this period increases, while the effect of vegetation alleviates.

3 Conclusions and discussions

With the validation of in situ data, the accuracy of the soil moisture data by the dual-channel retrieval algorithm and monthly regression analysis is better than that by official AMSR-E product, i.e., by which the soil moisture is accurately measured. However, due to the low spatial resolution of the passive microwave sensor, it cannot be applied in the retrieval of soil moisture in the middle and low-scales. In the subsequent measurement, the downscaling method that combines the high resolution data should be employed to acquire the soil moisture with more accurate and spatially higher.

In this study, we analyzed the temporal and spatial characterization of the soil moisture over the SRYR using the retrieval AMSR-E soil moisture products. Temporally, soil moisture is the highest in July and August and the lowest in January and December. Soil moisture generally tends to decrease in the spring, while increase again in the summer, have fluctuation in the autumn, and have little change in the winter. Spatially, the soil moisture is high in the southeastern section of the study area, and low in northwestern section; of which is the highest in the Zoige hummocky plateau, and lowest in the eastern margin of the Qaidam mountains. On the changing trend, the average soil moisture tends to decrease over the study period (2003-2010) at a rate of 0.012 cm³/(cm³·a), with the highest rate of decrease (0.067 cm³/(cm³·a)) in the Zoige hummocky plateau. The soil moisture in the wide valley basin of the YRSR and in the eastern margin of the Qaidam mountains increases slightly, i.e., the soil moisture decreases the most in the areas with initially high soil moisture, while increases in the areas that originally had low soil moisture.

Among all considered influencing factors, precipitation results in the greatest impact on soil moisture. Air temperature is negatively correlated with soil moisture during periods of high temperature, and positively during periods of low temperature. Actually, the climate in the SRYR shows a significant trend towards warmer and wetter during the study period^{[16, 29]}, while the soil moisture in the SRYR generally decreases, indicating that the increase of precipitation is not enough to offset the evaporative loss from warming trends and the amount of soil water available for vegetative growth. Previous study^[30] stated that when the air temperature increased by 2 ℃ and precipitation increased by less than 15%, evapotranspiration prevailed and the drought-related stress on the ecosystem increased; which could be eliminated only precipitation increased b 15% at the same time. Warming trend must, therefore, be offset by large increases in precipitation in order to ensure sufficient soil moisture and stabilize vegetative growth. In the next 90 years, air temperature in the QTP expectedly increases by 2.5 ℃^[31], while the increase of precipitation will not exceed 5% over the next 30-50 years^[32]. In another words, the increasing precipitation would not enough to compensate the negative effect of increasing air temperature, which would thereby intensify the drought trend in the SRYR.

The NDVI and soil moisture is positively correlated. Vegetation degradation further contributes to lowering soil moisture, especially significantly in the surface soil layers, and the highest reduction to the surface soil moisture was 38.6%^[28]. In the source region of 3 Rivers (Yangtze River, the Yellow River, and Lancang River) in the QTP, the ecological degradation of meadows also showed a strong impact on soil water conservation^[33], and thus, restoration and reconstruction of degraded ecosystems should play a positive effect on suppressing drought trend in the SRYR. Other than precipitation, air temperature, and NDVI, environmental factors such as topography, soil physical properties, and human activities cause certain impacts on soil moisture, and therefore they should be taken into the further research in the SRYR.