Soil Moisture Inversion in Grassland Ecosystem Using Remote Sensing Considering Different Grazing Intensities and Growing Seasons

Abstract

Although vegetation community information such as grazing gradient, biomass, and density have been well characterized in typical grassland communities with Stipa grandis and Leymus chinensis as dominant species, their impact on the soil moisture (SM) inversion is still unclear. This study investigated the characteristics of a grassland vegetation community at different grazing gradients and growing seasons and its impact on SM inversion using remote sensing data. The water cloud model (WCM) was used for SM inversion, and both field and remote sensing data collected from 2019 to 2021 were used for calibration and prediction. The study found that the calibrated WCM achieved prediction results of SM inversion with average R² values of 0.41 and 0.38 at different grazing gradients and growing seasons, respectively. Vegetation biomass and height were significantly correlated with vegetation indexes, and the highest model prediction accuracy was achieved for biomass and height around 121.1 g/m² [102.3–139.9] and 18.6 cm [17.3–19.8], respectively. Generally, NDWI₁ produced higher SM estimation accuracy than NDWI₂. The growing season of vegetation also affects the accuracy of the WCM to retrieve SM, with the highest accuracy achieved in mid-growing season I. Therefore, the developed WCM with optimal height and biomass of vegetation communities can enhance the SM prediction capacity; it thus can be potentially used for SM prediction in typical grasslands.

1. Introduction

Soil moisture (SM) plays an important role in the global water cycle [1] and vegetation growth [2,3], and its spatiotemporal dynamics control biomass accumulation during the growing season in steppe ecoregions [4,5]. Grassland vegetation growth exhibits a high sensitivity to SM variations [6]. The study by Fay et al. [7] demonstrates that soil moisture has a more pronounced impact on plant growth and biomass than on seed germination. SM is the critical factor in determining water stress and regulating interannual variability in grassland gross primary production, particularly in areas with shallow soil [8]. Furthermore, SM substantially affects grassland plant diversity [9]. Thus, comprehensive and accurate soil moisture data are essential for monitoring grassland drought and vegetation growth, monitoring grassland distribution, and estimating biomass [10].

Microwave remote sensing, both active and passive, has been successful in retrieving soil moisture over large areas due to its sensitivity to soil moisture, all-weather observation capability, low cost, and penetration capability through the vegetation canopy and atmospheric layers [11,12]. To date, microwave remote sensing of soil moisture has become one of the most useful technologies in retrieving soil moisture [13]. As a type of active microwave remote sensing, synthetic aperture radar (SAR) emits microwave radiation and measures the backscattered signal to obtain information about the target. It provides high spatiotemporal resolution and can penetrate clouds, rain, and other atmospheric interferences [14]. SAR data collected by Sentinel-1A (S-1A) and Sentinel-2 (S-2) provide essential soil information with high revisit frequency [15,16,17,18,19]. The backscattering coefficient of SAR to ground targets can be influenced by soil moisture, surface roughness, and vegetation characteristics [20,21].

The water cloud model (WCM) has been widely used in soil moisture retrieval from satellite remote sensing data, particularly in the case of SAR data [22,23,24]. This model [25] assumes that the surface roughness and vegetation properties are uniform at the pixel scale and that the soil moisture content varies horizontally and vertically according to a log-normal distribution. It also assumes that the soil and vegetation layers are horizontally stratified and that the microwave signal can penetrate through the vegetation layer but not the soil layer. The accuracy of SM estimation by the WCM can be impacted by many factors, mainly including regional characteristics [26,27], vegetation characteristics such as growth stage and density [12], as well as vegetation descriptors such as vegetation water content, leaf area index (LAI), normalized vegetation index (NDVI), normalized water index (NDWI), and enhanced vegetation index (EVI) [12,15,27,28,29]. Thus, it is extremely important to use appropriate vegetation descriptors according to specific regional characteristics and vegetation conditions to correct model parameters when using the WCM to estimate SM.

To date, most studies have primarily examined the influence of various vegetation types on SM inversion [30,31,32,33,34,35]. For example, Lei et al. [35] developed an enhanced WCM to retrieve the soil moisture in areas covered by greenwood, deciduous forest, mixed forest, composite shrub grass, and grassland. The distinct vegetation characteristics of these diverse vegetation types result in significant differences in WCM parameter values. However, relatively low prediction accuracy was obtained for all vegetation types examined. This outcome was attributed to the detrimental effect of the dense canopies, particularly greenwood, on the C-band microwave, underscoring the importance of vegetation index selection (e.g., NDVI and NDWI) for specific vegetation types. Researchers have further investigated the effect of seasonal variability among different vegetation types on SM estimates using the WCM [36,37]. In one study [37], the vertical-vertical (VV) polarization mode of S-1 C-band SAR images was used to retrieve soil moisture content (SMC), considering the characteristics of various crops and their conditions during different growth periods. The NDVI, extracted from S-2 MSI optical images, was combined with the WCM to eliminate vegetation moisture content (VMC) and obtain the surface backscattering coefficient, which was subsequently used to establish the SM model. Nonetheless, the mechanisms through which seasonal and regional differences in vegetation growth influence SM inversion remain unclear. Additionally, the influence of factors such as the vegetation communist structures (e.g., grazing gradient, biomass, density) and on the SM estimation through the WCM is not well understood.

The aim of this study was to evaluate the SM retrieval ability of the WCM combined with vegetation community information using Sentinel satellite data in a typical grassland. The study first characterized the soil and vegetation in a grassland under various grazing gradients and growing seasons from 2019 to 2021 by conducting field measurements. Then remote sensing data from Sentinel 1-A and Sentinel 2 were collected. The two types of data are used to calibrate the water cloud model (WCM) and evaluate the accuracy of soil moisture predictions based on different vegetation indexes (NDVI and NDWI). The study then relates spatiotemporal variations in vegetation community characteristics, such as height, biomass, density, and water content, to the vegetation indexes used to estimate soil moisture. The improved WCM and the optimization of key parameter coefficients are useful for improving SM prediction in grasslands.

2. Methods

2.1. Study Area

The corresponding technology roadmap is shown in Figure 1. The study area is located at Xilinhot National Climate Observatory in Inner Mongolia, China (44°08′–09′ N, 116°20′–21′ E, about 1127 m above sea level), as illustrated in Figure 2. This area is characterized by a typical continental climate with an average annual temperature of 2.5 °C and average annual rainfall of about 268 mm, supporting a temperate semi-arid grassland. It comprises a typical grassland, with a community composed mostly of xerophytic perennial bushy grasses and rhizome grasses. The dominant species are Stipa grandis, Cleistogenes squarrosa, Leymus chinensis, and Artemisia frigida.

The grazing gradients were established by block design and divided into four treatments with increasing grazing degree for comparison (Figure 3 and Figure 4): enclosure (NG, 0 sheep·ha⁻¹), light grazing (LG, 2 sheep·ha⁻¹), moderate grazing (MG, 4 sheep·ha⁻¹), and heavy grazing (HG, 8 sheep·ha⁻¹). The grazing gradients and their grading refer to the study [38]. To obtain the representative data at different locations within the study zone with the four grazing gradients, the study zone was divided into five blocks of five units each. Each block had an area of 120 m × 120 m (Figure 3). The blocks that met the criteria of each grazing gradient were selected as the study blocks. To avoid the sampling distance being smaller than the minimum distance that can be captured by the satellite (i.e., pixel size of 10 m × 10 m), the sampling distances between two sampling points in a block in 2019 and 2020–2021 are greater than 20 m. The study period was from 28 July 2019 to the end of 9 October 2021, which covers different growing seasons of the vegetation, including early growing season (EGS), mid-growing season I (MGS-I), and mid-growing season II (MGS-II).

2.2. Measurement of Basic Characteristics of Soils and Vegetation

The real-time kinematic (RTK) means that a fixed base station and a mobile receiver were used to provide highly accurate positioning data in real time. Before field sampling, we measured the latitude and longitude to determine the sampling location. Sampling was conducted two days after the S-1A acquisition. The basic characteristics of plant species at each sampling point were measured using the quadrat sampling method. Three 1 m × 1 m quadrats were set along each transect in the study area. Species names, heights, and the number of individuals were recorded in each quadrat. All aboveground vegetation was clipped, brought back to the laboratory, and weighed after drying at 65 °C. This dry weight refers to the vegetation biomass of each sampling block. The vegetation community height in each quadrat was determined by taking the average height of all vegetation. The plant community density was represented by the average number of individuals per quadrat. The vegetation water content (VWC) was calculated by subtracting the dry weight from the fresh weight of plant samples and then dividing the result by the area of the sampling block. The mean vegetation water content for each plot was determined by averaging the values from three 1 m × 1 m quadrates. At each sample point in the study area, soil samples with depths of 0–10 cm were collected by ring knife method (Figure 3), after which the dried soil bulk density (ρ), soil mass water content (SM_m), and volumetric water content (SM) were calculated as follows:

$$\rho =\frac{{\mathrm{S}}_{\mathrm{d}}}{{\mathrm{V}}_{\mathrm{c}}}$$

(1)

$${\mathrm{S}\mathrm{M}}_{\mathrm{m}}=\frac{{\mathrm{S}}_{\mathrm{w}}-{\mathrm{S}}_{\mathrm{d}}}{{\mathrm{S}}_{\mathrm{d}}}$$

(2)

$$\mathrm{S}\mathrm{M}=\rho \times {\mathrm{S}\mathrm{M}}_{\mathrm{m}}$$

(3)

where ρ represents the soil bulk density (g/cm⁻³), V_c represents the ring knife volume (cm³), and S_w and S_d represent the wet and dry weights of soils, respectively.

A total of 900 valid samples were obtained, containing all grazing gradients and including different growing seasons and different years, of which 720 samples were randomly selected to calibrate the model parameters and the remaining 180 samples were used to evaluate the model prediction accuracy. This sampling method and sample treatment are the same as the following two sections of data gathering.

2.3. Remote Sensing Data from Sentinel-1A

As a radar imaging satellite, the Sentinel-1A satellite operates in the C-band frequency and uses a synthetic aperture radar (SAR, 5.405 GHz, 5.6 cm wavelength) instrument to capture images of the Earth’s surface. Sentinel-1A was launched by the European Space Agency (ESA) in April 2014. The revisit cycle of Sentinel-1A is 12 days with two radar modes, i.e., interferometric wide swath (IW) and extra wide swath (EW). The IW ground range detected with height (GRDH) data with the polarization of VV were downloaded from the ASF platform (https://vertex.daac.asf.alaska.edu/ (accessed on 5 January 2022)) and used in this study. Note that the pixel size is 10 m × 10 m in the IW-GRDH mode, and Sentinel-1A with the VV polarization has been widely used for soil moisture retrieval [39].

Restricted by the relatively long revisit cycle of Sentinel-1A (i.e., 12 days), we collected 13 S-1A images of the study site from 28 July 2019 to 9 October 2021. These images were extracted from the satellite database according to the geographic coordinates of the sampling location. They were captured with an incident angle of approximately 36.6°. All the images were processed using the SNAP software developed by the ESA (http://step.esa.int/main/toolboxes/snap (accessed on 10 January 2022)), which includes orbital correction, radiometric calibration, filtering, terrain correction, and conversion of intensity values to the vegetation-soil backscatter coefficient. The refined LEE (local equilibrium equation) filter was used for speckle correction. Range Doppler correction was used for the geometrical correction of the images with the aid of SRTM 3 s DEM.

2.4. Remote Sensing Data from Sentinel-2A

The Sentinel-2 multi-spectral instrument (MSI) is an imaging instrument on board the Sentinel-2 satellites, which are part of the European Union’s Copernicus Earth observation program. The MSI can capture images of the Earth’s surface in 13 spectral bands, ranging from the visible to the near-infrared, with a ground sampling distance of 10–60 m. The Sentinel-2 mission includes two identical satellites, Sentinel-2A and Sentinel-2B, which were designed to work together to provide complete and continuous coverage of the Earth’s surface for land monitoring applications, such as vegetation monitoring. They were launched by the ESA in June 2015 and March 2017, respectively. In this study, we selected Level-2A products that were processed and calibrated Sentinel-2 satellite images that have been corrected for atmospheric effects and provide surface reflectance values. These images were extracted from the satellite database at the imaging date close to that of Sentinel-1A. A total of 13 acquired images were selected and analyzed using SNAP software for subsequent vegetation index.

2.5. Soil Moisture Model and Parameter Inversion

The WCM partitions the total backscattering coefficient (${\sigma }_{\mathrm{t}\mathrm{o}\mathrm{t}}^0$) of a vegetation-covered surface into two components: volume scattering directly reflected from the vegetation layer (${\sigma }_{\mathrm{v}\mathrm{e}\mathrm{g}}^0$) and direct backscattering from the soil surface (${\sigma }_{\mathrm{s}\mathrm{o}\mathrm{i}\mathrm{l}}^0$) after two attenuations. The volume scattering component (${\sigma }_{\mathrm{v}\mathrm{e}\mathrm{g}}^0$) is related to vegetation characteristics such as biomass, density, and vegetation water content. This component mainly contributes to the low-frequency part of the radar signal and is sensitive to the structure and growth status of the vegetation. The soil backscattering component (${\sigma }_{\mathrm{s}\mathrm{o}\mathrm{i}\mathrm{l}}^0$) is associated with soil moisture content and surface roughness, and it dominates the high-frequency part of the radar signal. This component is primarily affected by the dielectric properties and roughness of the soil surface. Their relationship and calculation expressions are as follows:

$${\sigma }_{\mathrm{tot}}^0\left(\theta \right)={\sigma }_{\mathrm{veg}}^0\left(\theta \right)+{\tau }^2\left(\theta \right){\sigma }_{\mathrm{soil}}^0\left(\theta \right)$$

(4)

$${\sigma }_{\mathrm{veg}}^0\left(\theta \right)=\mathrm{a}{\mathrm{V}\mathrm{W}\mathrm{C}}_1\mathrm{c}\mathrm{o}\mathrm{s}\left(\theta \right)\left[1-{\tau }^2\left(\theta \right)\right]$$

(5)

$${\tau }^2\left(\theta \right)=\mathrm{e}\mathrm{x}\mathrm{p}\left[-2\mathrm{b}{\mathrm{V}\mathrm{W}\mathrm{C}}_2/\mathrm{c}\mathrm{o}\mathrm{s}\left(\theta \right)\right]$$

(6)

where ${\tau }^2\left(\theta \right)$ is the two-way attenuation factor (transmittance) of radar waves penetrating vegetation layer, and θ is the radar incidence angle; the constant terms of a and b depend on the vegetation types, and radar configuration, respectively; VWC₁ and VWC₂ are vegetation parameters used as descriptors of the canopy. They can be represented by bulk vegetation water content (VWC) [40].

In practice, it is difficult to measure the VWC values in the field. Thus, they are usually estimated by the normalized difference vegetation index (NDVI) or normalized difference water index (NDWI) [41]. Note that the NDVI is calculated from the reflectance of visible and near-infrared light, while the NDWI is calculated from the reflectance of the near-infrared and short-wave infrared bands. These two indexes (NDVI and NDWI) are then selected as parameters to characterize VWC in this study. The empirical conversion formula of VWC with the NDVI or NDWI for grazed land can be expressed as [42,43]:

$$\mathrm{V}\mathrm{W}\mathrm{C}=4.2857\times \mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}-1.5429 (\mathrm{NDVI}>0.5)$$

(7)

$$\mathrm{V}\mathrm{W}\mathrm{C}=1.9134\times \mathrm{N}\mathrm{D}\mathrm{V}{\mathrm{I}}^2-0.3215\times \mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I} (0.17<\mathrm{NDVI}\le 0.5)$$

(8)

$$\mathrm{V}\mathrm{W}\mathrm{C}=0 (\mathrm{NDVI}\le 0.17)$$

(9)

$$\mathrm{V}\mathrm{W}\mathrm{C}=1.44\times {\mathrm{N}\mathrm{D}\mathrm{W}\mathrm{I}}^2+1.36\times \mathrm{N}\mathrm{D}\mathrm{W}\mathrm{I}+0.34$$

(10)

Studies have shown the potential of using short-wave infrared (SWIR) and narrow-edge NIR measurements in enhancing the calculations of the NDWI [44]. In order to better characterize changes in the vegetation canopy, we enhance the NDWI using short-wave infrared (SWIR) and narrow-edge NIR of Sentinel-2, which is sensitive to vegetation changes. The expressions for the NDVI and NDWI are as follows:

$$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}=\left({\mathrm{R}}_{\mathrm{N}\mathrm{I}\mathrm{R}}-{\mathrm{R}}_{\mathrm{R}\mathrm{E}\mathrm{D}}\right)/\left({\mathrm{R}}_{\mathrm{N}\mathrm{I}\mathrm{R}}+{\mathrm{R}}_{\mathrm{R}\mathrm{E}\mathrm{D}}\right)$$

(11)

$${\mathrm{N}\mathrm{D}\mathrm{W}\mathrm{I}}_1=\left({\mathrm{R}}_{\mathrm{n}\mathrm{a}\mathrm{r}\mathrm{r}\mathrm{o}\mathrm{w}-\mathrm{e}\mathrm{d}\mathrm{g}\mathrm{e}\mathrm{N}\mathrm{I}\mathrm{R}}-{\mathrm{R}}_{\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}\right)/\left({\mathrm{R}}_{\mathrm{n}\mathrm{a}\mathrm{r}\mathrm{r}\mathrm{o}\mathrm{w}-\mathrm{e}\mathrm{d}\mathrm{g}\mathrm{e}\mathrm{N}\mathrm{I}\mathrm{R}}+{\mathrm{R}}_{\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}\right)$$

(12)

$${\mathrm{N}\mathrm{D}\mathrm{W}\mathrm{I}}_2=\left({\mathrm{R}}_{\mathrm{N}\mathrm{I}\mathrm{R}}-{\mathrm{R}}_{\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}\right)/\left({\mathrm{R}}_{\mathrm{N}\mathrm{I}\mathrm{R}}+{\mathrm{R}}_{\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}\right)$$

(13)

where ${\mathrm{R}}_{\mathrm{R}\mathrm{E}\mathrm{D}}$, ${\mathrm{R}}_{\mathrm{N}\mathrm{I}\mathrm{R}}$, ${\mathrm{R}}_{\mathrm{n}\mathrm{a}\mathrm{r}\mathrm{r}\mathrm{o}\mathrm{w}-\mathrm{e}\mathrm{d}\mathrm{g}\mathrm{e}\mathrm{N}\mathrm{I}\mathrm{R}}$, and ${\mathrm{R}}_{\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}$ represent the reflectance of band 4 (665 nm), band 8 (842 nm), band 8A (865 nm), and band 11 (1610 nm) of S-2, respectively.

Generally, the relationship between backscattering intensity and soil moisture is not linear but rather exhibits a saturation effect. As soil moisture increases, the backscattering signal initially increases linearly but then reaches a maximum point beyond which further increases in soil moisture do not result in a corresponding increase in backscattering intensity. This is due to the attenuation of the microwave signal as it passes through the soil, which reduces the amount of signal that is returned to the satellite. Given that our soil moisture has a range of 0–0.274 cm³/cm³ with a mean value of 0.133 cm³/cm³ and a standard error of 0.002 cm³/cm³, the relationship between backscattering intensity and soil moisture can be modeled using a linear function. This simplification has also been used by previous studies [40,45,46]. The relationship between backscattering intensity and soil moisture is written as follows:

$${\sigma }_{\mathrm{soil}}^0\left(\theta \right)=\mathrm{c}·\mathrm{S}\mathrm{M}+\mathrm{d}$$

(14)

The backscattering coefficients can thus be expressed as follows by combining Equations (4)–(6) and (14):

$$\mathrm{S}\mathrm{M}=\frac{\left[{\sigma }_{\mathrm{tot}}^0\left(\theta \right)-\mathrm{a}\times \mathrm{V}\mathrm{W}\mathrm{C}\mathrm{c}\mathrm{o}\mathrm{s}(\theta )·\left(1-\mathrm{e}\mathrm{x}\mathrm{p}\left(-\frac{2\mathrm{b}\times \mathrm{V}\mathrm{W}\mathrm{C}}{\mathrm{c}\mathrm{o}\mathrm{s}\left(\theta \right)}\right)\right)\right]/\mathrm{e}\mathrm{x}\mathrm{p}\left[-\frac{2\mathrm{b}\times \mathrm{V}\mathrm{W}\mathrm{C}}{\cos \left(\theta \right)}\right]-\mathrm{d}}{\mathrm{c}}$$

(15)

Using Equations (7)–(15), the surface soil water content after removing the vegetation effect can be calculated. The model coefficients were solved using the least squares method with input parameters including SM, backscatter coefficient, VWC, and incidence angle. We used 80% of the field measurement data and remote sensing data for the parameter inversion. These data were randomly selected from the whole data package.

2.6. Accuracy Evaluation of Soil Inversion Model

The accuracy of the model is evaluated using the remaining 20% of the field measurement data and remote sensing data. The accuracy of the model is evaluated by calculating the determination coefficient (R²) and root mean square error (RMSE) using the observed and estimated values of SM. They are commonly used as model accuracy evaluation indexes in the research community of ecology.

The calculation formulas for R² and RMSE are shown below.

$${\mathrm{R}}^2=1-\frac{\sum _{i=1}^n {\left(M_i-P_i\right)}^2}{\sum _{i=1}^n {(M_i-\overline{M})}^2}$$

(16)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{\frac{1}{n}{\sum }_{i=1}^n{\left(M_i-P_i\right)}^2}$$

(17)

where M_i represents the observed SM, P_i represents the predicted SM of the model, and $\overline{M}$ represents the average of the observed SM.

3. Results

3.1. Field Measurement Results

The measured vegetation community information at the four-point scale of grazing gradients (NG, LG, MG, and HG) in the investigated grassland ecosystem is shown in Figure 5. As expected, the vegetation height, biomass, and density generally decreased significantly with increased grazing intensity. However, the volumetric water content initially decreased with higher gradient intensity but stabilized afterward. Although the biomass analysis found that the biomass at grazing gradient MG was slightly higher than that at LG, there was no significant difference between them (Figure 5b). The vegetation density at grazing gradient MG is lower than that at HG, indicating that the density is not only influenced by grazing gradient but also other factors, such as physical and chemical properties of soil, disturbances from precipitation, as well as plant–herbivore interactions.

The measured vegetation characteristics at the three stages of vegetation growing seasons from the early stage to the later middle stage (ESG, MGS-I, and MGS-II) are also shown in Figure 6. Generally, the vegetation biomass and the density generally increased significantly with the growth of vegetation. However, both vegetation height and soil volumetric water content initially decreased and then increased with the growth of vegetation. This fluctuation indicates the complexity of vegetation growth and its interaction with soils during different growing seasons.

The soil moisture and standard error calculated from the field samples are presented with different grazing gradients from low to high grazing intensity in Figure 7a. The maximum value of mean soil moisture is 0.148 kg/m² without grazing treatment (i.e., the NG group), while the minimum value of mean soil moisture is 0.123 kg/m² when light grazing is performed (i.e., the LG group). In addition, similar to the variation trend of the volumetric water content of vegetation with the grazing gradient, the soil moisture content decreased hugely when the grazing gradient increased to light grazing from ungrazed level then fluctuated with higher grazing intensity up to heavy grazing intensity. Figure 7b further shows the measured soil moisture at different growing seasons. The soil moisture substantially increases with the growth of vegetation and finally tends to be stable at the later middle growing season.

3.2. Inversion Results of WCM Parameters

The coefficients of WCM parameters (i.e., a, b, c, and d) were inverted using 80% remote sensing data and 80% field measurement data as inputs.

Table 1. The WCM parameter inversion results for the various grazing gradients.

Grazing Gradients	Vegetation Index	a	b	c	d	R2	RMSE
NG	NDVI	−11.0	−0.3	64.6	−22.5	0.46	0.028
	NDWI1	−29.9	−0.8	44.1	−18.6	0.37	0.027
	NDWI2	−29.2	−0.4	54.9	−21.7	0.36	0.027
LG	NDVI	−114.7	2.7	89.1	−34.1	0.61	0.031
	NDWI1	−23.3	1.0	65.0	−28.6	0.59	0.031
	NDWI2	86.6	−12.1	−60.2	14.5	0.69	0.029
MG	NDVI	−19.7	−0.3	74.8	−25.4	0.41	0.026
	NDWI1	−44.2	−1.0	49.4	−19.8	0.40	0.026
	NDWI2	−208.7	0.0	79.1	−27.1	0.38	0.026
HG	NDVI	−18.8	−3.3	134.7	−27.4	0.32	0.022
	NDWI1	−20.3	−1.6	118.8	−24.8	0.06	0.012
	NDWI2	−5.4	−14.1	−9.4	0.3	0.09	0.013

Note: NG: enclosure; LG: light grazing; MG: moderate grazing; HG: heavy grazing. NDVI: the normalized difference vegetation index; NDWI₁/NDWI₂: two normalized water indexes. a, b, c, and d are coefficients of the WCM. R²: the determination coefficient. RMSE: root mean square error.

and

Table 2. The WCM parameter inversion results for various growing seasons.

Growing Seasons	Vegetation Index	a	b	c	d	R2	RMSE
Early growing season (EGS)	NDVI	−17.1	−24.9	−48.7	1.5	0.56	0.020
	NDWI1	0.0	−19.5	−0.1	0.0	0.11	0.014
	NDWI2	−31.1	−3.1	−40.7	−5.9	0.05	0.009
Mid-growing season Ⅰ (MGS-I)	NDVI	−13.7	−0.2	72.2	−25.0	0.47	0.027
	NDWI1	−41.2	−1.0	49.9	−19.6	0.49	0.027
	NDWI2	−50.0	−0.3	64.1	−23.7	0.45	0.027
Mid-growing season Ⅱ (MGS-II)	NDVI	−13.8	−0.3	74.6	−25.2	0.40	0.027
	NDWI1	−29.1	−1.1	51.9	−19.0	0.44	0.027
	NDWI2	8.3	−0.3	69.5	−24.0	0.39	0.027

Note: NDVI: the normalized difference vegetation index; NDWI₁/NDWI₂: two normalized water indexes. a, b, c, and d are coefficients of the WCM. R²: the determination coefficient. RMSE: root mean square error.

show the model parameter inversion results for the various grazing gradients and growing seasons, considering that different vegetation indexes (NDVI, NDWI₁, and NDWI₂) are used for calculating VWC and thus SM. The R² and RMSE values are presented to evaluate the performance of the calibrated model.

The model was first calibrated for the vegetation under grazing gradients with increasing intensity (NG, LG, MG, and HG). Table 1 presents the performance metrics of the model, with an average R² value of 0.40 [0.32–0.59] and a standard error of 0.054. The model also shows an average RMSE of 0.025 [0.022–0.029] with a standard error of 0.002. The values in the square brackets represent the lower 95% confidence interval of the mean and the upper 95% confidence interval of the mean. The low RMSE values suggest that the model calibration is robust under different grazing gradients. The coefficients derived under various grazing gradients can be used for subsequent model accuracy testing. Furthermore, the one-way ANOVA results indicate that there is no significant difference in the mean values of RMSE among the NDVI, NDWI₁, and NDWI₂ groups (p = 0.33). The mean values of coefficients a, b, c, and d are −36.55 (20.170), −2.52 (1.490), 58.74 (15.110), and −19.60 (3.933), respectively, as illustrated in Figure 8a.

The study also involved calibrating the model for vegetation at different growing seasons, ranging from early to later stages. Table 2 displays the average R² value of the model at 0.37 [0.110−0.490], with a standard error of 0.058. The RMSE has an average value of 0.023 [0.014−0.027], with a standard error of 0.002. The low RMSE values indicate a well-calibrated model under different grazing gradients. The one-way ANOVA analysis shows no significant differences in the mean values of RMSE among the NDVI, NDWI₁, and NDWI₂ groups (p = 0.85). The mean values of coefficients a, b, c, and d are −20.86 (6.276), −5.663 (3.117), 35.52 (16.42), and −15.66 (3.678), respectively, as shown in Figure 8. Interestingly, the rank sequence among a, b, c, and d (i.e., c > b > d > a) is similar to that obtained when calibrating the model for vegetation under different grazing gradients (Figure 8).

3.3. Model Accuracy Evaluation

We substituted 20% of the field measurement data and 20% of the remote sensing data into the WCM with the determined coefficients to evaluate the accuracy of the WCM. The evaluation indicators are R² and RMSE.

Table 3. The WCM accuracy evaluation results for various grazing gradients.

Grazing Gradient	Vegetation Index	R2	Average R2	RMSE	Average RMSE
NG	NDVI	0.32	0.23	0.033	0.032
	NDWI1	0.19		0.031
	NDWI2	0.18		0.031
LG	NDVI	0.63	0.68	0.024	0.022
	NDWI1	0.66		0.020
	NDWI2	0.74		0.022
MG	NDVI	0.43	0.38	0.025	0.026
	NDWI1	0.36		0.026
	NDWI2	0.34		0.027
HG	NDVI	0.53	0.34	0.020	0.016
	NDWI1	0.28		0.012
	NDWI2	0.20		0.015

Note: NG: enclosure; LG: light grazing; MG: moderate grazing; HG: heavy grazing. NDVI: the normalized difference vegetation index; NDWI₁/NDWI₂: 2 normalized water indexes. R²: the determination coefficient. RMSE: root mean square error.

shows the results of WCM accuracy evaluation under increasing grazing gradients. The average value of R² between all the actual values and the predicted values of the model is 0.41 [0.20−0.63] and a standard error of 0.056. The corresponding average value of RMSE is 0.024 [0.020−0.031] with a standard error of 0.002.

By averaging the evaluation index values of three vegetation indexes, the average R² and RMSE for each grazing gradient are also presented. A relatively high average R² of 0.68 is observed for grazing gradient LG; a low average RMSE of 0.022 is also found. Meanwhile, for the grazing gradient LG, the vegetation index NDWI₂ resulted in more accurate SM retrieval than the other two vegetation indexes (R² = 0.74, RMSE = 0.022). The model regression result between the estimated soil moisture and the observed one for LG using NDWI₂ is illustrated in Figure 9a. For the grazing gradients NG, MG, and HG, the vegetation index NDVI resulted in more accurate SM retrieval than the other two vegetation indexes. For example, for HG, the R² value of 0.53 is obtained by using the NDVI, much higher than those obtained by using NDWI₁ and NDWI₂. Therefore, the NDVI is more suitable for establishing the soil moisture retrieval model (i.e., WCM) for the grazing gradients NG, MG, and HG, while NDWI₂ is more suitable for the WCM when the grazing gradient is LG. It was also found that under most grazing conditions, NDWI₁ generally produced higher estimation accuracy than NDWI₂; thus, better applicability can be achieved.

Table 4. The WCM accuracy evaluation results for various growing seasons.

Growing Seasons	Vegetation Index	R2	Average R2	RMSE	Average RMSE
Early growing season (EGS)	NDVI	0.44	0.25	0.033	0.035
	NDWI1	0.15		0.030
	NDWI2	0.15		0.043
Mid-growing season Ⅰ (MGS-I)	NDVI	0.50	0.51	0.026	0.026
	NDWI1	0.52		0.026
	NDWI2	0.50		0.025
Mid-growing season Ⅱ (MGS-II)	NDVI	0.35	0.39	0.029	0.028
	NDWI1	0.45		0.028
	NDWI2	0.36		0.028

Note: NDVI: the normalized difference vegetation index; NDWI₁/NDWI₂: two normalized water indexes. R²: the determination coefficient. RMSE: root mean square error.

shows the results of WCM accuracy evaluation under various growing seasons. The average value of R² between all the actual values and the predicted value of the model is 0.38 [0.15−0.50] with a standard error of 0.048. The corresponding average value of RMSE is 0.030 [0.026−0.033] with a standard error of 0.002. These results show that the model in this study has high accuracy under various growing seasons, and the prediction results of the model can be used for subsequent research.

By averaging the evaluation index values of three vegetation indexes, the average R² and RMSE for each growing season are also presented. A relatively high average R² of 0.51 is observed for MGS-I with a low average RMSE of 0.026. Thus, the WCM of MGS-I has the highest accuracy (R² = 0.51, RMSE = 0.026). Compared with the other two vegetation indexes, the precision of the model calculated by NDWI₁ is slightly better (R² = 0.52, RMSE = 0.026). The model regression result between the estimated soil moisture and the observed one for MGS-1 using NDWI₁ is illustrated in Figure 9b. For MGS-II, the WCM calculated by NDWI₁ has higher accuracy compared to the other two vegetation indexes. For EGS, the WCM calculated by the NDVI has higher accuracy. Therefore, in EGS, it is more accurate to estimate SM using the model with the NDVI as a vegetation index. In MGS, it is more accurate to estimate SM using the model with the NDWI₁ vegetation index. Generally, the WCM using NDWI₁ is more accurate for SM estimation in different growing seasons compared to that using NDWI₂.

3.4. Correlation between Vegetation Index and Vegetation Community Information

The vegetation indexes (NDVI and NDWI) that are used for estimating VWC and SM in this paper are calculated from the measured spectral reflectance data reflected by vegetation at certain wavelength intervals. Thus, these vegetation indexes may be related to the measured vegetation community information, including vegetation height, density, biomass, and VWC. In this section, we further present a correlation analysis between them (Figure 10). The results show that biomass is significantly correlated with the vegetation indexes NDVI, NDWI₁, and NDWI₂ with coefficients of 0.82, 0.57, and 0.53, respectively. Additionally, the NDVI is significantly correlated with vegetation height (p < 0.01). Evidently, the primary factors influencing vegetation indexes in the context of plant communities are plant height and biomass.

4. Discussion

In this study, a comprehensive investigation of the vegetation community information (i.e., vegetation height, biomass, density, and volumetric water content) at three stages of vegetation seasons shows that they have rather distinct characteristics, which produce different optimized model parameter coefficients (a, b, c, and d) with different model training accuracies. Consequently, the SM prediction capacities with different vegetation indexes (NDVI and NDWI) are influenced by the vegetation community information (i.e., grazing gradient, biomass, and density).

Given that the parameter calibration of WCM is critical for the inversion of SM, both field measurement data and remote sensing data are inputted for calibration, causing acceptable training results with average R² values of 0.40 and 0.37 at different grazing gradients and growing seasons, respectively. The RMSE values are 0.025 and 0.023, respectively. Note that the variation of the measured vegetation community information with increasing grazing gradient intensity is consistent with other studies [47], indicating the accurate measurements of vegetation communities. Regarding the evaluation results of SM after model training, slightly higher R² values and lower RMSE values are obtained compared to the training values, as shown in Table 3 and Table 4. For example, the average value of R² is 0.41 and 0.38 at different grazing gradients and growing seasons, respectively. This prediction accuracy of the soil moisture inversion model is generally comparable to other studies where the WCM is used for SM prediction. For example, by using NDVI and NDWI as vegetation indexes, Tucker [48], Chen et al. [49], and Liang et al. [37] applied WCM to evaluate the SM, producing R² (RMSE) values of 0.31 (0.299), 0.36 (0.267), and 0.51 (3.725), respectively. It should be noted that the spectral characteristics of grassland vegetation are complex, which may lead to the relatively low accuracy of the model for retrieving SM. The use of detailed grassland classification methods may refine estimates of SM in grassland. Lei et al. [35] have shown that other vegetation types including greenwood and mixed forest could have a higher surface soil moisture.

The WCM shows that the SM estimation is determined by the accurate estimation of VWC, which is calculated by different vegetation indexes such as the NDVI and NDWI (Equations (7)–(10)). By doing so, the correlation between the vegetation index and vegetation community information can be used for linking the SM estimation accuracy to the vegetation community information. In our study, the model prediction accuracy under grazing gradient LG is the highest for all vegetation indexes (Table 3). Figure 10 shows that biomass is significantly correlated with the vegetation indexes NDVI, NDWI₁, and NDWI₂. The correlation coefficient between the vegetation indexes and biomass ranged from 0.53 to 0.82 (p < 0.001). This significant correlation agrees well with the study from Zhang and Wang [50], where the correlation coefficient ranged from 0.71 to 0.76 (p < 0.01). In addition, height is significantly correlated with the NDVI, though the correlation coefficient of 0.47 is not high. As a result, the highest model prediction accuracy can be achieved for biomass around 121.1 g [102.3–139.9] and for height around 18.6 cm [17.3–19.8], as shown in Figure 5a,b. In practice, vegetation biomass may be potentially used as a biomarker for applying the WCM to the SM prediction in a typical steppe. It is noted that the dual-polarized radar vegetation index [51], dual-polarization SAR vegetation index (DPSVI) [52], and dual-polarimetric radar vegetation index (DpRVI) [23,53] have been used in crop condition monitoring.

The growing season of vegetation could also affect the accuracy of WCM to retrieve SM. In MGS-I, the WCM has the highest accuracy (Table 4), indicating that MGS-I may be the most suitable to mitigate the scattering effect of vegetation on the backscattering coefficient of microwave signals in remote sensing for improving the model prediction capacity. The backscattering coefficient of microwave signals from the Earth’s surface could be influenced by the dielectric properties of vegetation, the orientation of vegetation, the shape and size of vegetation elements, and the density and moisture content of vegetation [54,55]. Therefore, understanding the scattering effect of vegetation is crucial in accurately retrieving parameters such as soil moisture from remote sensing data. Further studies can be carried out to systematically investigate the scattering effect of vegetation with different community structures on the backscattering coefficient of microwave signals in remote sensing and the SM prediction capacity of the WCM. The robustness of SM prediction capacity at different years of vegetation growth can also be estimated.

5. Conclusions

This study presents a modified water cloud model (WCM) with optimized model parameters, developed using field data and Sentinel satellite data. The model’s training and testing performances are impacted by grassland vegetation community information at different grazing gradients throughout different growing seasons. This community information also affects the selection of vegetation indexes (NDVI and NDWI) for soil moisture (SM) prediction. Our findings indicate that the vegetation height, biomass, and density generally decreased substantially with increased grazing intensity, while the biomass and density increased significantly throughout the growing seasons.

Parameter calibration of the water cloud model (WCM) is essential for SM inversion, necessitating both field measurement data and remote sensing data for calibration. The study yields acceptable prediction results, with average R² values of 0.41 and 0.38 for different grazing gradients and growing seasons, respectively. The prediction accuracy of the SM inversion model is generally comparable to other studies employing WCM for SM prediction. Notably, vegetation biomass and height exhibit significant correlations with vegetation indexes. The optimal model prediction accuracy was attained for biomass and height at approximately 121.1 g/m² [102.3–139.9] and 18.6 cm [17.3–19.8], respectively. Moreover, NDWI₁ generally outperforms NDWI₂ in terms of estimation accuracy in most instances. The vegetation growing season also impacts the accuracy of WCM in retrieving SM, with the highest accuracy achieved in the mid-growing season I. Taking into account the community structure and growing season of grassland vegetation, which are influenced by factors such as grazing, offers a theoretical foundation for the estimation of soil moisture in grassland ecosystems.

Comprehending the scattering effect of vegetation on the backscattering coefficient of microwave signals in remote sensing is crucial for accurately retrieving parameters such as SM. Further studies can explore this effect with diverse community structures and investigate the influence of the vegetation growth year on SM prediction.