Assessment of High-Resolution Surface Soil Moisture Products over the Qinghai–Tibet Plateau for 2009–2017

Abstract

The surface soil moisture over the Qinghai–Tibet Plateau (QTP) has an important impact on the weather and climate of East Asia. Under climate warming, the imbalance of solid–liquid water of QTP has become a research hotspot, but the surface soil moisture dynamics over QTP are not clear owing to the lack of precise measurements over a large scale. In this paper, the quality of gridded surface soil moisture products including CSSPv2 high-resolution (6 km) simulation, ESA CCI satellite retrieval, ERA5 land-atmosphere coupled reanalysis, and GLDAS2.1 land reanalysis products (Noah, Catchment, VIC) is analyzed over QTP by comparison with the in situ measurements at 140 stations during 2009–2017. We find that the CSSPv2 product shows a higher correlation than the global satellite and reanalysis products, with correlation increased by 7.7%–115.6%. The root mean squared error of the CSSPv2 product is lower than that of other products, with the error decreased by 13.4%–46.3%. The triple collocation analysis using high-resolution simulation, global reanalysis, and satellite retrieval products over the entire plateau shows that the error of CSSPv2 is the lowest, followed by ESA CCI, while ERA5 is the highest. The soil moisture products of ESA CCI, ERA5, and CSSPv2 all show an increasing trend from April to September of 2009 to 2017, with wetting in the west and drying in the east. This study indicates that the CSSPv2 high-resolution surface soil moisture product has better performance over QTP than other global products, and the global satellite and reanalysis products may overestimate the surface soil moisture dynamics.

1. Introduction

Soil moisture refers to the water content stored in the soil, which is an important element in hydrology, meteorology, and other related fields [1,2]. Research shows that the role of soil moisture in climate change is secondary only to sea surface temperature (SST), and its role even exceeds that of SST over land [3]. Because of the thermal and hydraulic characteristics of various physical processes in soil, the soil moisture changes slower than the atmosphere, causing an anomaly that can affect the subsequent weather and climate conditions [4].

The soil moisture over the Qinghai–Tibet Plateau (QTP), which is the largest plateau in China and the highest plateau in the world, is sensitive to climate change. With climate warming, the solid–liquid water of QTP is out of balance, and the increase in liquid water leads to the instability of the “Asian water tower”. Some studies pointed out that persistently subtle changes in hydrological and thermal processes in the QTP may have an important impact on China’s weather and climate, as well as even on the atmospheric circulation in Asia and the world through dynamic and thermodynamic processes [5,6,7,8,9,10]. Thus, it is of great significance to study the change in soil moisture in QTP to cope with climate change and adaptive utilization of water resources. However, precise measurement of soil moisture is challenging over QTP, which hinders further investigation.

At present, the methods of obtaining soil moisture data mainly include in situ measurement, satellite remote sensing, and land surface model simulation. The accuracy of in-situ observation is the highest. However, because of the complex terrain, harsh climate, and sparse population of QTP, there are few soil moisture stations. Currently, the observational stations are mainly distributed in Naqu, Maqu, Pali, and Ngari [11,12,13,14]. Compared with site observation, the model and satellite data can provide continuous soil moisture products in time and space, as well as cost less resources. The commonly used land model products include GLDAS data from NASA and ERA5 data from the European Center for Medium-Range Weather Forecasts (ECMWF). The commonly used satellite data include ESA CCI data from the European Center, SMAP data from NASA, and so on [11,15,16,17,18,19,20,21]. However, because of the inconsistency in the evaluation data and regimes, the verifications result in different conclusions in different studies. For example, Xing et al. [17] evaluated seven sets of satellite and model products, such as GLDAS-Noah, ESA CCI, and ERA5, by using the site observation in permafrost regions of QTP, and found that ESA CCI products had a higher correlation and lower root mean square error (RMSE). Deng et al. [15] evaluated four sets of GLDAS data (Noah, CLM, VIC, and MOSAIC) using site observation, and found that Noah model products have good performance. Bao et al. [19] evaluated the GLDAS-Noah data using observations from four networks over the QTP, and found that the GLDAS-Noah soil moisture biases mainly show up in errors of values in the nonfrozen period.

Although many studies have evaluated the soil moisture products in QTP, there are still shortcomings. QTP is an area with a complex terrain and uneven vegetation distribution, and the soil moisture distribution is highly heterogeneous. Previous studies are mostly based on a single observation network, the verifications do not take into account the spatial heterogeneity, and the spatial resolution of the global products is relatively coarse (e.g., 30 km). With the continuous development of computing resources and satellite data, the spatial resolution of soil moisture products can be increased, which provides a favorable tool for studying soil moisture dynamics at a fine scale. For example, Ji et al. [22] used a high-resolution land surface model (CSSPv2) to make land surface hydrothermal products with a resolution of 6 km in China, and revealed that CSSPv2 high-resolution soil moisture products were more accurate than ERA5, GLDAS-Noah, ESA CCI, and other products by verification at more than 1000 soil moisture observational stations. However, because of the limitation of observation data over QTP, it is necessary to assess the high-resolution products with several soil moisture observational networks at highlands.

Besides using traditional metrics, including correlation coefficient and RMSE, that can be applied at station scale with in situ observations, Stoffelend [23] developed the triple-collocation (TC) method, which can avoid seeking the “true value”. This method can objectively estimate the error standard deviation of three independent datasets. This method has been widely used in the verification of soil moisture [24,25,26,27]. Moreover, the TC method is applicable to various spatial scales, which makes up for the uneven spatial distribution of observation stations.

This study will assess several sets of high-resolution soil moisture products in QTP using the data of nearly 140 observational stations. Section 2 introduces soil moisture products such as in situ observations, satellite, reanalysis, and high-resolution simulation products, as well as the methods for evaluating datasets. Section 3 presents the results, and Section 4 presents the conclusions and discussion.

2. Data and Method

2.1. In Situ Observations

The observational data used in this study are from 140 observation stations of the National Tibetan Plateau Data Center (http://data.tpdc.ac.cn (accessed on 16 August 2021)) [28,29,30,31,32,33]. The sites’ distribution is shown in Figure 1. These stations are mainly located in four different areas, namely, the Ngari observation network in a cold arid area, Naqu observation network in a cold semi-arid area, Maqu observation network in a cold humid area, and Pali observation network. The depth of observation data is 0–5 cm, the study period is from 2009 to 2017, and the temporal resolution is processed to daily scale.

2.2. GLDAS-2.1

The goal of the NASA Global Land Data Assimilation System (GLDAS) is acquiring satellite and ground observation data to generate the best land surface state and flux field using an advanced land surface model and data assimilation technology [34]. GLDAS-2.1 uses observation data from 2000 to the present as the forcing data. In this study, the simulation products of Noah, VIC, and Catchment in GLDAS-2.1 are used. The depth of data is 0–10 cm and the spatial resolution is 0.25 degrees.

2.3. ESA CCI

The Soil Moisture CCI project is part of the ESA Programme on Global Monitoring of Essential Climate Variables (ECV), better known as the Climate Change Initiative (CCI), initiated in 2010, which has produced an updated soil moisture product every year [35,36,37]. ESA CCI soil moisture products have contributed to hundreds of hydrological and climatic studies all over the world. ESA CCI soil moisture products are composed of three surface soil moisture data sets: active, passive, and combined [38]. The “active product” and “passive product” are produced by fusing the soil moisture products of scatter and radiometer, respectively; the “combination product” is a hybrid product based on the “active product” and “passive product”. In this study, the ESA CCI “combined product” (V06.1) is used. The coverage of the dataset is over 40 years, from November 1978 to 31 December 2020.

2.4. ERA5

ERA5 is the fifth generation ECMWF reanalysis of global climate and weather [39]. Reanalysis is to combine model data with observation data from all over the world by using the laws of physics to form a global complete and consistent dataset. ERA5 offers many hourly estimates of the atmosphere, waves, and land surface states. ERA5 has four layers of soil moisture, with depths of 7 cm, 21 cm, 72 cm, and 189 cm, respectively. The depth of 0–7 cm is selected in this study. The spatial resolution is 0.25 degrees.

2.5. CSSPv2

The second generation Conjunctive Surface-Subsurface Process model (CSSPv2) used in this study is based on the Common Land Model [40,41], with a substantial improvement in hydrological processes [18,22,42,43,44,45]. The runoff generation scheme is storage-based and the hydrological influence of soil organic matter is added to CSSPv2 to improve its performance in simulating land hydrology in an alpine basin [43]. The model has been applied in several mountainous areas and it shows good performance in hydrological variables such as soil moisture, runoff, evapotranspiration, and total water storage [42,43,46]. The CSSPv2 model is driven by CLDAS meteorological data with a spatial resolution of 6 km. The CSSPv2 model is calculated in 11 soil layers; we selected the surface soil and interpolated it to 0–10 cm. As the site probe only monitors the change in soil water content, only the soil water is considered when verifying the products, and soil ice is added when analyzing the change characteristics.

2.6. Evaluation Metrics and Evaluation Method

We use traditional evaluation metrics in this research, including the correlation coefficient (CC), root mean square error (RMSE), and bias.

The correlation coefficient indicates the fitting degree between soil moisture products and the in situ observations, which can be calculated as follows:

$$CC=\frac{{{\displaystyle \sum }}_{t=1}^N\left({\theta }_s\left(t\right)-{\mu }_s\right)\left({\theta }_m\left(t\right)-{\mu }_m\right)}{\left(N-1\right){\delta }_s{\delta }_m}$$

(1)

Root mean squared error means the absolute deviation and accuracy of soil moisture products when compared with the in situ observations, and the calculation method is shown below:

$$RMSE=\sqrt{\frac{{{\displaystyle \sum }}_{t=1}^N{\left({\theta }_s\left(t\right)-{\theta }_m\left(t\right)\right)}^2}{N}}$$

(2)

Bias means the systematic difference between soil moisture products and the in situ observation, which can be calculated by Formula (3):

$$bias=\frac{{{\displaystyle \sum }}_{t=1}^N\left({\theta }_s\left(t\right)-{\theta }_m\left(t\right)\right)}{N}$$

(3)

where θ_s is the simulated soil moisture and θ_m is the in situ observation. In time series analysis, N is the overall time range and t represents a specific time. μ_s and μ_m are the mean simulated and observed values during the overall time range, respectively, while δ_s and δ_m represent the standard deviations.

Considering the temporal coverage of each product and the influence of soil freezing and thawing, the evaluation period is from April to September of 2009 to 2017. In order to obtain more accurate results, this study takes the grid of GLDAS as the standard and all the soil moisture products were interpolated to the GLDAS 0.25° grid. Then, these grids containing observational stations were determined. Note, for the grid that covers more than one station, the mean values of different stations within the same grid were calculated and used to perform the evaluation. All of the evaluation metrics were first calculated over each grid containing observational stations and then averaged over QTP.

2.7. Triple-Collocation (TC) Method

The TC method can be used to estimate the random error variance of three groups of datasets describing the same physical quantity, and it is an effective method to evaluate soil moisture products [47]. The TC method has four assumptions: ① every dataset is linearly correlated with the true value; ② the error is stable and does not change with time; ③ errors of datasets are independent of each other; and ④ the error of the dataset is independent of the truth value. Thus, the following estimates can be obtained from assumption ①:

$${\theta }_i=a_i\theta +b_i+{\epsilon }_i,$$

(4)

where ${\theta }_i^{}$ and ${\epsilon }_i^{}$ represent observation data of the ith (i ∈ { 1,2,3}) dataset θ_i and its corresponding random error, respectively, and each contains N observations of the same geophysical process. θ is the true value. According to [24], the TC method can be solved using the covariance method.

$$Cov\left({\theta }_i,{\theta }_j\right)=a_i{a}_j{\sigma }_{\theta }^2+a_{i}Cov\left(\theta ,{\epsilon }_i\right)+a_{j}Cov\left(\theta ,{\epsilon }_j\right)+Cov\left({\epsilon }_i,{\epsilon }_j\right)$$

(5)

where ${\sigma }_{\theta }^2$ is the variance of the truth data set θ. According to hypotheses ③ and ④, Equation (5) can be simplified into Equation (6):

$$C_{ij}=Cov\left({\theta }_i,{\theta }_j\right)=\left\{\begin{array}{c}{a}_i{a}_j{\sigma }_{\theta }^2,i\ne j\\ a_i{a}_j{\sigma }_{\theta }^2+{\sigma }_{{\epsilon }_i}^2,i=j\end{array} \right.$$

(6)

Furthermore, the estimation expression of the error standard deviation is as follows:

$${\sigma }_{{\epsilon }_i}=\left\{\begin{array}{c}\sqrt{C_{11}-\frac{C_{12}{C}_{13}}{C_{23}}},i=1\\ \sqrt{C_{22}-\frac{C_{12}{C}_{23}}{C_{13}}},i=2\\ \sqrt{C_{33}-\frac{C_{13}{C}_{23}}{C_{12}}},i=3\end{array}\right.$$

(7)

3. Results

3.1. Evaluation against In Situ Observations

Figure 2 shows that CSSPv2 has the highest correlation with the observations during 2009–2017, with an average correlation of 0.61. ERA5 has the secondary highest correlation, with an average value of 0.56. The correlations of other products from high to low are ESA, Noah, Catchment, and VIC, and the corresponding average values are 0.52, 0.45, 0.39, and 0.28, respectively. The average correlation coefficients of CSSPv2 are 7.7%, 16.3%, 35.9%, 56.5%, and 115.6% higher than those of the ERA5, ESA, Noah, Catchment, and VIC products, respectively. For the root mean square error (RMSE), CSSPv2 has the lowest, with an average value of 0.083 m³/m³. The RMSEs from low to high are 0.096 m³/m³, 0.101 m³/m³, 0.108 m³/m³, 0.132 m³/m³, and 0.155 m³/m³ for ESA, Noah, Catchment, ERA5, and VIC, respectively. Compared with products such as ESA, Noah, Catchment, ERA5, and VIC, the average RMSE of CSSPv2 is 13.4%, 17.7%, 22.9%, 37.0%, and 46.3% smaller, respectively. For the bias, only CSSPv2 showed an average dry deviation, with an average of −0.017 m³/m³, and other products showed an average wet deviation. Noah showed the smallest deviation of 0.006 m³/m³, and the deviations of other products from low to high were ESA, Catchment, ERA5, and VIC, with average values of 0.016 m³/m³, 0.038 m³/m³, 0.088 m³/m³, and 0.120 m³/m³, respectively.

Figure 3 shows the evaluation results of each soil moisture product in four sub-regions. Although the performances of each product in different regions are different, CSSPv2 still has the best performance.

It can be seen from Figure 4 that the variations in soil moisture over the Maqu area are well simulated by CSSPv2 and ERA5, while the simulation results of VIC in early April are almost opposite to the observation. In the Naqu area, the CSSPv2 and ESA products have the smallest deviation. In the Ngari area, the CSSPv2 product is the closest to the observation, while ERA5 underestimates the soil moisture and other products generally overestimate the soil moisture. Yuan et al. [43] compared the GLDAS-2.1 products and CSSPv2 high-resolution simulation over the Maqu and Naqu networks, and found that using the high-resolution meteorological forcings that fuse many more observations (and thus have higher accuracy) and the improvement in model parameterizations (e.g., considering the hydraulic influence of soil organic matters) both contribute to better performance of the CSSPv2 simulation. In addition, CSSPv2 uses the state-of-the-art land surface dataset, which also helps to improve the soil moisture simulation [22]. For example, the 1 km resolution soil texture dataset used by the CSSPv2 model can provide more accurate hydraulic information (e.g., soil porosity and saturated hydraulic conductivity), which then influences the soil moisture simulation.

3.2. Triple-Collocation Analysis

In the evaluation using in situ observation, several products with good performance can be found, such as CSSPv2, ESA, ERA5, and Noah. Among them, the CSSPv2 and ESA products perform well in terms of correlation coefficient, RMSE, and deviation. Considering the independence assumptions of the TC method, three sets of data, CSSPv2, ESA, and ERA5, are selected in the calculation of the TC error in this study. Because the spatial resolutions of ESA, ERA5, and CSSPv2 are different, the data of ESA and ERA5 are interpolated to 6 km as CSSPv2.

Firstly, we analyze the correlation among three sets of soil moisture data. Figure 5 shows the correlation coefficients between different products for the daily soil moisture climatology and daily soil moisture anomaly. It can be seen that three products show a positive correlation in most areas of the plateau, except for some areas in the northwest and edge of the plateau. The correlation coefficient between ERA5 and CSSPv2 is higher for the climatology fields than the anomaly fields, suggesting that the two products are more consistent in representing seasonal cycles of soil moisture. However, ESA is less correlated with ERA5 and CSSPv2 even with seasonal cycles, because the mean correlation coefficients are much smaller than that of ERA5/CSSPv2 (Figure 5a,c). For the soil moisture anomaly, the three datasets are more correlated with each other over eastern QTP than western QTP. In general, the average correlation coefficient of ESA and ERA5 is the highest (Figure 5d), while correlations of CSSPv2 and ESA and of ESA and CSSPv2 are similar (Figure 5e,f).

It is noted that the three datasets are less correlated over the northwestern part of QTP (masked out by black rectangles in Figure 5). Figure 6 further presents the time series of soil moisture averaged over northwestern QTP. The ESA soil moisture shows a large variation over this region, which may be unrealistic because of the satellite observation noise [35]. The large variation in ESA leads to a lower and even negative correlation between the ESA and CSSPv2 or ERA5 datasets. CSSPv2 and ERA5 also show some inconsistent soil moisture variations in some years, leading to a low correlation coefficient. For example, ERA5 shows a decreasing trend of soil moisture in 2009, while CSSPv2 shows an increasing trend. ERA5 soil moisture increases first and then decreases in 2017, while CSSPv2 shows less fluctuations. Different soil moisture variation between ERA5 and CSSPv2 may be related to the use of different precipitation [22]. As few meteorological stations are located over northwestern QTP, little observational information can be used in the generation of both station-based (used by CSSPv2) and reanalysis (used by ERA5) precipitation products, leading to large discrepancies.

Figure 7 shows the spatial distribution of the standard error deviation of ESA, ERA5, and CSSPv2 soil moisture in the study area. In order to ensure the reliability of the TC results, the standard deviation of the error was calculated only for the areas where the correlation coefficient between the three groups of data was greater than 0.2 and passed the significance test (significance α = 0.05). On the spatial average, the uncertainty of ERA5 is the largest, followed by ESA, and the uncertainty of CSSPv2 is the lowest. From the spatial distribution, the precision of the ESA product is higher in the central, eastern, and southeast part of the plateau, but lower in the western and northeast part of the plateau. The uncertainty of the ERA5 product is larger in the western and northern parts of the plateau, but smaller in the eastern part of the plateau. CSSPv2 products are different to the former two products; the uncertainty of CSSPv2 products is larger in the middle and southeast parts of the plateau, and smaller in the west part of the plateau. According to the uncertainty distribution of the three products, it can provide a reference for the usage and improvement of the products in QTP. For example, CSSPv2 products can be used in the western part of the plateau with high accuracy and ESA products can be used in the eastern part of the plateau with high accuracy. It should be noted that the conclusions obtained are only limited to the places with TC results, thus the areas not covered need further discussion.

3.3. Analysis of Spatial and Temporal Changes

The soil moisture of ESA shows a significant drying trend mainly in the east-central and southwest marginal areas of the plateau, as well as a significant wetting trend in the west-central regions (Figure 8a). ERA5 soil moisture shows an increasing trend in the west and decreasing trend in the east (Figure 8b). The CSSPv2 soil moisture product shows a decreasing trend in the central and northeast parts of the plateau, but the values are scattered and smaller than those of the ESA and ERA5 products (Figure 8c). CSSPv2 shows a significant increasing trend in the western plateau. In summary, ESA, ERA5, and CSSPv2 products show a wetting trend in the west part of QTP and a drying trend in the east part of QTP from April to September of 2009 to 2017. Given that CSSPv2 has the highest accuracy, we speculate that the drying trend in the eastern part might be overestimated (Figure 8).

Figure 9 presents the regional mean soil moisture anomalies. It can be found that all three soil moisture products show an increasing trend, among which CSSPv2 has the fastest increasing rate, which is about 0.018 m³/m³/decade, and ERA5 and ESA have similar growth rates, which are about 0.01 m³/m³/decade. These data have three periods of change in 2009–2017, which are the increase in 2009–2012, the reduction in 2012–2015, and the increase in 2015–2017, and the growth rate in 2015–2017 is faster than that in 2009–2012. The ERA5 soil moisture product has larger variability than the other two products.

4. Conclusions and Discussion

Based on the observation data of 140 stations in the Qinghai–Tibet Plateau (QTP), we assessed several soil moisture products, including a high-resolution land surface modeling product, the satellite products, and global reanalysis products, from April to September in 2009–2017. The main conclusions of this study are as follows.

(1) As validated against in situ observations, CSSPv2, ESA, and ERA5 products show a higher correlation coefficient; CSSPv2, Noah, and ESA products have a lower root mean square error; while CSSPv2, Noah, and ESA products have a smaller deviation. Among the six products, the CSSPv2 high-resolution land surface modeling product has the best performance. The better performance of CSSPv2 is attributed to the higher accuracy of the meteorological forcings and advanced model parameterizations.

(2) Based on the triple-collocation (TC) analysis, it is found that the uncertainty of CSSPv2 is the smallest, that of ESA is in the middle, and that of ERA5 is the largest. The uncertainty of ESA and ERA5 is larger in the western plateau and smaller in the eastern plateau, but the uncertainty of CSSPv2 is smaller in the western plateau and larger in the eastern plateau.

(3) The soil moisture products of ESA, ERA5, and CSSPv2 show an increasing trend from April to September of 2009 to 2017, with CSSPv2 showing the fastest growth rate. All three products show a wetting trend in the west and drying trend in the east of the plateau.

Our verification results are consistent with those of Zeng et al. [18], where CSSPv2 shows better simulation than ERA5 and GLDAS/Noah. The evaluation results of other products are also consistent with previous studies [15,17]. For the TC analysis, using three different sets of data will obtain different results, but the error distributions are basically unchanged. For the error distribution of these three sets of products, CSSPv2 is low in the west and high in the east, while ERA5 and ESA are high in the west and low in the east, which is also consistent with the TC error distribution. Thus, the high-resolution soil moisture products (e.g., CLDAS/CSSPv2) can provide intermediate proxy data for climate studies, which show superiority over existing global reanalysis products and satellite products over QTP.

Although two evaluation methods are used in this paper to comprehensively assess soil moisture products in QTP, there are still some deficiencies in the research. For example, the amount of improvement in the CSSPv2 simulation against GLDAS-2.1 and ERA5 products attributed to the meteorological forcings or model structures is not quantified here. Future efforts are needed to resolve this issue by conducting a series of experiments (e.g., driving different models with different meteorological forcings). In addition, the observation stations are densely located in the four areas, and only 32 grids contain the 140 in situ stations. Although these four observational networks are typical, continued efforts and new networks are needed to figure out the performance of different products over other regions (e.g., northwestern QTP) because of the highly heterogeneous environment in QTP.