Collaborative Inversion of Soil Water Content in Alpine Meadow Area Based on Multitemporal Polarimetric SAR and Optical Remote Sensing Data

Soil water content is a critical environmental parameter in research and practice, though various technological and contextual constraints limit its estimation in arid areas with vegetation cover. Tis study combined the multitemporal remote sensing data of Sentinel-1 and Landsat 8 to conduct an inversion study on surface soil water content under low vegetation cover in Nagqu, central Tibetan Plateau. Four vegetation indices (NDVI, ARVI, EVI, and RVI) were extracted from optical remote sensing data. A water cloud model was used to eliminate the infuence of the vegetation layer on the backscattering coefcient associated with vegetation cover, and a predictive model suitable for the Nagqu area was constructed. Te water cloud model efectively incorporated a vegetation index instead of vegetation water content. We found that VV polarization was more suitable for soil water content inversion than VH polarization. Among the four vegetation indices, the soil water content inversion model constructed with RVI under VV polarization had the best ft ( R 2 � 0.8212; RMSE � 6.30). Te second-best ft was observed for vegetation water content-NDVI ( R 2 � 0.8201). Te soil water content inversion models all had an R 2 > 0.6, regardless of the vegetation index used, though the RVI had the best ftting efect, indicating that this vegetation index is highly applicable in the water cloud model, as a substitute for vegetation water content, and is expected to perform well in similar study sites.


Introduction
Soil water content (soil moisture) is not only a basic condition for plant growth but is also an important environmental parameter in the felds of ecology, hydrology, agriculture, and climate change and a useful indicator for drought monitoring and crop yield estimation, among others.It is one of the most important parameters for characterizing ground surfaces [1,2].Terefore, large-scale monitoring of soil water content is not only an important part of agricultural research and the evaluation of environmental factors but also of great importance for developing regional and global climate mitigation strategies and predicting regional dry and wet conditions.
Traditional soil water content monitoring is mainly conducted using established specialized monitoring or meteorological stations.Although this method has high accuracy, it is limited by a small detection range, low turnover, and data representativeness and requires considerable resources [3][4][5].With the maturation of satellite remote sensing technology, multisource remote sensing can generate large-scale soil water content estimates with high spatial and temporal resolution, facilitating the acquisition of dynamic real-time information and compensating for the shortcomings of traditional monitoring methods [6].
Techniques for soil water content inversion using remote sensing are mainly categorized into optical remote sensing inversion, microwave remote sensing inversion, and collaborative inversion [7][8][9].Optical remote sensing inversion is performed by extracting sensitive information related to soil water content from optical remote sensing images.According to the diferent extraction approaches, it can be divided into the refectivity, index, and thermal inertia methods [10].Optical remote sensing cannot directly provide spectral information for soil in densely vegetated areas and is mainly applicable to the inversion of bare surfaces.Compared with optical remote sensing, microwave remote sensing, categorized as active or passive according to the working method of the sensor, has the advantages of not being afected by weather, strong penetration, and allweather capabilities [11][12][13].Passive microwave remote sensing inverts the soil water content using a prediction model established with soil radiation temperature (obtained using the microwave radiometer) and measured soil data [14][15][16].Active microwave remote sensing, which is less afected by interference, is based on the mathematical relationship between the backscattering coefcient and soil water content, among which the linear regression model is the most commonly used [17][18][19].Te radar backscattering coefcient of active microwave remote sensing is mainly afected by the soil dielectric constant, surface roughness, and vegetation cover.In the vegetation-covered area, the scattering and attenuation efect of the vegetation layer on the radar backscatter echo will reduce the sensitivity of the backscatter coefcients to the change in soil water content, thus afecting the accuracy of the inversion.Terefore, the efect of the vegetation cover on the backscatter coefcients needs to be eliminated.
Vegetation index can provide information about vegetation growth status and cover, so vegetation index can be used instead of vegetation cover.At the same time, the relationship between diferent types of vegetation index and soil water content may be diferent.Terefore, understanding the applicability of vegetation index is helpful to better select the soil water content inversion model suitable for specifc regions and environments, so as to improve the accuracy and reliability of soil water content inversion.Alternatively, the sensitivity of optical remote sensing to vegetation information can be used to extract relevant vegetation canopy information, and the infuence of the vegetation layer can be eliminated by water cloud [20] or MIMICS models [21], which enable collaborative inversion with microwave remote sensing and improve the inversion accuracy of soil moisture in vegetation-covered areas.For example, Liu et al. [22] combined Sentinel-1A and -2A images, used the water cloud model to remove the infuence of the vegetation layer, and then estimated soil moisture in the farmland area using machine learning.Rawat et al. [23] estimated soil moisture by combining the soil's dielectric constant properties with the radar backscattering coefcient obtained using the SA schemes of RISAT-1 data.Su and Cao [24] used full-polarization data to invert soil moisture using AIEM, water cloud, and MIMICS models according to land use types and vegetation in diferent rocky desert areas.
Our study locality, Nagqu City, Tibet, is located at high altitudes with a cold and dry climate, making drought an important factor afecting agricultural production.An efective inversion method can improve the monitoring of soil water content and enable the dissemination of this crucial information to producers at all levels within this region.In this study, multitemporal remote sensing images with a large time span were used for collaborative inversion of soil water content.Given the unique context of this study locality, we explored the applicability of the vegetation index instead of vegetation water content to further improve drought detection capabilities in this region. .

Study Area.
Te Nagqu soil temperature and humidity observation network in the central Qinghai-Tibet Plateau is located within a 100 × 100 km area (91.5 °-92.5 °E, 31 °-32 °N, Figure 1) at an average altitude of 4650 m above sea level.Most of the selected sites were located in the central and northern parts of Nagqu County, and a few were located southeast of Amdo County and southwest of Nyainrong County.It belongs to the high-altitude subcold climate zone, where the total annual hours of sunshine are 2852.6-2881.7,annual average temperature is −0.9 to −3.3 °C, annual average relative humidity is 48-51%, and the annual average precipitation is 380 mm.Most of the study area is located in the pure summer and autumn pastoral area of Nagqu.Te vegetation type is a typical alpine meadow belt, mainly composed of wormwood meadows.Te grass layer has a height of 1-5 cm and total annual growth days of 150-170 d, indicating that the withering period is longer than the growing period and vegetation cover is generally low.

Remote Sensing Data.
Sentinel-1, a follow-up satellite of ERS-2 and Envisat, is an important component of the European Space Agency's Copernicus Program (previously known as GMES) and integral to the Earth observation program.Te constellation comprises Sentinel-1A and Sentinel-1B satellites equipped with a C-band synthetic aperture radar (SAR) at an orbital altitude of 693 km.Te observation period of a single satellite is 12 d, and the combined observation of two satellites is 6 d.Sentinel-1 has four imaging modes, SMStripmap (SM), interferometric wide swath (IW), extrawide swath (EW), and wave mode (WV).Te GRD data in the IW mode were selected for this study.Te obtained SAR image comprised L1 data with a spatial resolution of 5 × 20 m, with VV and VH dual polarization (Table 1).Te downloaded image data were subjected to orbit correction, thermal noise removal, radiometric calibration, multiview, fltering, terrain correction, and decibel processing in the ESA SNAP.We used Refned Lee fltering to obtain the total backscattering coefcient of the radar.
Landsat 8 is an Earth observation satellite jointly launched by the National Aeronautics and Space Administration (NASA) and the United States Geological Survey (USGS) on February 11, 2013, carrying two sensors, a land imager (OLI) and a thermal infrared sensor (TIRS).Te OLI  Journal of Spectroscopy includes nine bands and the TIRS includes two thermal infrared bands.Optical images from the same day as the SAR data were downloaded from the USGS website (https:// earthexplorer.usgs.gov).Landsat 8 L2 SP T1-level OLI data were used to calculate the normalized diference vegetation index (NDVI), enhanced vegetation index (EVI), ratio vegetation index (RVI), and atmospheric impedance index (ARVI) in ENVI 5.6 to obtain the corresponding grid image (Table 2).
To promote the accuracy of the model, we selected optical and microwave remote sensing images for the same date; the cloud content of the optical remote sensing images was 5%.

Soil Water Content Data.
Te soil water content data were obtained from the multiscale observation network of soil temperature and moisture on the central Tibetan Plateau provided by the National Tibetan Plateau Data Center [25,26], which mainly considers soil water content in the surface layer (0-5 cm) at 57 monitoring stations.By screening and eliminating abnormal data from six image datasets, our fnal dataset considers data for 178 monitoring points for the construction and verifcation of the remote sensing image inversion model.

Analytical Approaches
2.3.1.Quantifcation of Soil Water Content.Soil water content is usually expressed as soil weight water content, soil volumetric water content, and relative soil water content.Soil weight water content refers to the weight ratio of soil water to dried soil.Te soil volumetric water content refers to the volumetric ratio of soil water per unit of soil.Relative soil water content refers to the percentage of soil water content relative to feld capacity.
Te soil water content data obtained in this study refected soil weight water content; therefore, we converted the data to soil volumetric water content using the following formula: where m v is the soil volumetric water content (%), m q is the soil weight water content (%), and ρ is the soil bulk weight (g•cm −3 ).

Optical and Microwave Remote Sensing Collaborative
Inversion Method.We observed a linear correlation between the soil backscattering coefcient and soil water content, according to which the soil water content inversion model can be established as follows: where m v is the soil volumetric water content (%), δ 0 soil is the soil backscattering coefcient (dB), and a and b are the empirical parameters.

Water Cloud Model.
Te scattering and attenuation efects of the vegetation layer cause partial refection of the signal received by the sensor.To eliminate the infuence of the vegetation layer on the inversion of soil water content in areas with vegetation cover, we applied the water cloud model proposed by Baghdadi et al. [18], which is often used to invert soil water content in areas with low vegetation cover [27,28].Te water cloud model divides the total received radar backscatter coefcient (δ 0 total ) into two parts, the soil scattering coefcient through the vegetation (δ 0 soil ) and the vegetation scattering coefcient (δ 0 veg ), and is modeled as Te collation can be deduced as In equations ( 3)-( 6), δ 0 total is the total radar backward scattering coefcient (dB), δ 0 veg is the vegetation layer backward scattering coefcient (dB), δ 0 soil is the soil backward scattering coefcient (dB), c 2 is the two-time attenuation factor of the vegetation scattering coefcient (%), VWC is the vegetation water content (g•cm −3 ), θ is the radar incidence angle ( °), and A and B are the empirical vegetation parameters.As the study area is dominated by alpine meadows, the empirical parameters A and B were set to 0.0014 and 0.084, respectively, according to Bindish and Barros [29] (Table 3).

Estimation of Vegetation Water Content.
Vegetation water content is an important parameter in water cloud models and can be inverted using optical remote sensing data based on three methods: spectral refectivity, spectral index, and radiative transfer model methods.Te spectral index method has a higher accuracy than the refectivity method and is simpler than the radiative transfer method; therefore, it is more widely used [30].
Te NDVI and normalized diference water index (NDWI) are commonly used to estimate vegetation water content using water cloud models.To explore the feasibility of diferent vegetation indices for characterizing the VWC in the water cloud model, we frst calculated a reference vegetation water content estimate using the empirical relationship between the NDVI and VWC established by Jackson and Gao et al. [31,32].With the formula, VWC NDVI � 0.098 × e 4.225NDVI . (7) Second, the EVI, RVI, and atmospherically resistant vegetation index (ARVI) were directly substituted into the vegetation water content parameter in the water cloud model to invert the soil water content.Compared to the VWC NDVI inversion, these vegetation indices were calculated as follows:

Accuracy Evaluation of Model.
We evaluated the accuracy of the inversion model of soil water content using goodness of ft (R 2 ) and root mean square error (RMSE), which can be calculated as follows: where ŷi is the inversion value of soil water content (%), y is the mean soil water content (%), y i is the measured value of soil water content (%), and n is the number of samples used in verifcation.R 2 represents the ftting relationship between the inversion and measured values; the closer it is to 1, the better the ft of the model.RMSE represents the degree of deviation between the inversion and measured values; the smaller the value, the better the ft of the model.
In this study, the vegetation index and backscattering coefcient in satellite images of six time-phases are extracted, respectively, and the soil backscattering coefcient is solved by using the water cloud model to eliminate the infuence of vegetation cover.Tere are a total of 178 available monitoring point data, each time-phase of about 30 data.Ten, the empirical parameters a and b are solved by linear ftting the measured data of some monitoring points with the soil backscattering coefcient.Finally, the remaining measured data are brought into the constructed inversion model to detect its accuracy.
Te technical workfow for inverting soil water content is shown in Figure 2.

Analysis of Soil Water Content Collaborative Inversion
Results.Previous studies reported a correlation between the radar backscattering coefcient and soil water content [33].Simultaneously, diferent polarization modes are afected by vegetation cover to varying degrees, resulting in diferent polarization modes and soil water content correlations [34,35].
According to the above principles and processes, data from 100 monitoring points in the reference dataset were selected to construct the inversion model, and the remaining 78 data points were used to verify the accuracy of the model.Te ratio of validation set to test set allocation is approximately 3 : 4. Firstly, the vegetation index and backscatter coefcient of 178 monitoring points were calculated by Journal of Spectroscopy    optical image and SAR, respectively, and the soil backscattering coefcient is extracted by using the water cloud model to eliminate the infuence of vegetation cover.Ten, the solved soil backscattering coefcient is linearly ftted with 100 measured data.Te ftting results of 100 measured soil water content and soil backscattering coefcient under VV/VH dual polarization are shown in Figure 3. Te ftting degree of vegetation water content and vegetation index on the soil backscattering coefcient and measured soil water content under VH polarization was lower than that under VV polarization (Figure 3).

Verifcation of Soil Water Content Inversion Accuracy.
Te accuracy of the inversion model was verifed using the dataset composed of the remaining 78 monitoring points based on goodness of ft (R 2 ) and root mean square error (RMSE).Te soil backscattering coefcients of 78 monitoring points were substituted into the above four linear models to invert the soil water content.Ten, the inversion value of soil water content is linearly ftted with the measured value.We observed a good ft between the backscattering coefcients and soil water content after removing the infuence of vegetation using VV and VH dual polarization (Figure 4), which is consistent with previous results [36].Consistent with our fndings in Section 3.1, we observed a better ft under VV polarization than under VH polarization.Under VV polarization, the RMSE did not vary much between models with diferent vegetation indices (the maximum was 6.84 for EVI and the minimum was 6.30 for RVI); however, we observed a large diference in R 2 (the maximum value was 0.61 for RVI and the minimum value was 0.55 for VWC-NDVI and ARVI).Considering the degree of ftting and the need for accurate prediction of the inversion value, the linear regression model y � −1.0924x -17.611 based on RVI in Figure 3(d) has the best predictive performance for soil water content estimation and is most suitable for the study area.

Conclusions
In this study, the NDVI, ARVI, EVI, and RVI were calculated using multitemporal data from Sentinel-1 GRD and Landsat 8. Te infuence of the vegetation layer was eliminated by using a water cloud model.Based on the model ft associated with the diferent vegetation indices, we constructed a predictive model for soil water content inversion in vegetation-covered areas and made the following conclusions: (1) Using Sentinel-1 and Landsat 8 remote sensing data for collaborative inversion, we found that the model ft and accuracy were better under VV polarization than those under VH polarization, proving that VV polarization was more suitable for soil moisture inversion in this area.(2) ARVI, EVI, and RVI were selected to replace the vegetation water content parameter in the inversion model, and all had R 2 values >0.6, indicating a good ftting efect.RVI performed better than VWC-NDVI (R 2 � 0.6744), suggesting that it can efectively replace vegetation water content in the water cloud model.
(3) By calculating NDVI, ARVI, EVI, and RVI and using the corresponding empirical relationship formula to convert NDVI to VWC-NDVI as a reference, we found that using diferent vegetation parameters to invert or replace vegetation water content can improve the water cloud model performance and enhance the accuracy of soil water content inversion to diferent extents.
We efectively applied NDVI, ARVI, EVI, and RVI to the collaborative inversion of soil water content.RVI had the best ftting efect, which may be related to the hilly and mountainous surface characteristics of the study area in Nagqu.We therefore propose that RVI is especially applicable in mountainous study sites, where it can eliminate various terrain efects to improve inversion accuracy.
Note that there are some limitations to our study.Te data provided by Sentinel-1 are in the VH dual polarization mode, while research on the HH polarization mode is lacking.Terefore, future studies should analyze the applicability of the remaining two polarization modes to obtain better inversion results.Although the vegetation layer and soil roughness greatly afect the radar backscattering coefcient, we only considered the efect of scattering from the vegetation layer on the inversion of soil water content.

8
Journal of Spectroscopy Terefore, subsequent studies could beneft from applying soil roughness to the inversion model, which may further improve model accuracy at our and comparable study sites.

Figure 1 :
Figure 1: Location of the study area (a), distribution of monitoring points (b), and vegetation types (c).

Table 1 :
Sentinel-1 SAR images acquired for the study area.

Table 2 :
Landsat 8images acquired for the study area.

Table 3 :
Parameter values of A and B under diferent vegetation types.
Tis indicates that the ability of microwaves to penetrate the vegetation layer under VV polarization is stronger than that under VH polarization, which is consistent with previous results.Meanwhile, the accuracy of the model constructed varied with diferent parameters for vegetation water content.Under VV polarization, the soil backscattering coefcients based on VWC-NDVI and RVI best ft with the measured soil water content data (R 2 � 0.6727 and 0.6744, respectively).EVI corresponded with the second-best ft (R 2 � 0.6323), while ARVI corresponded with the worst ft (R 2 � 0.6066).Under VH polarization, the soil backscattering coefcient and soil water content based on VWC-NDVI had the best ftting efect (R 2 � 0.2401), while ARVI had the worst ftting efect (R 2 � 0.1624).Te analysis showed that the ftting efect was better with the RVI than with the vegetation water content estimated by VWC-NDVI, indicating that RVI can directly replace the vegetation water content parameter in the model.Te goodness of ft associated with EVI and ARVI was also above R 2 > 0.6, indicating good model ft and applicability.