1. Introduction
Wheat is an important global food crop and accurate yield information is essential to achieving food security. The water content of wheat crops is an important growth indicator during different growth stages. It not only affects wheat photosynthesis, but also the grain filling rate and, ultimately, yield [
1,
2,
3]. Therefore, monitoring wheat crop water content is important for achieving rapid and accurate estimates of wheat yield during growth.
Traditionally, wheat crop water content has been determined by manually sampling plants to obtain fresh weight and dry weight, and then calculating crop water content. However, this method is time consuming, labor intensive, inefficient, and difficult or impossible to implement in large land areas [
4]. In recent years, the rapid development of remote sensing technology has produced a large amount of researches that applied remote sensing data for monitoring the vegetation water content in a large area, quickly and accurately [
5,
6,
7]. These studies have focused primarily on two categories of data sources: optical remote sensing and radar remote sensing.
Traditional optical remote sensing methods are mainly based on vegetation indices and radiation transmission models. Vegetation indices used to invert vegetation water content are based on water sensitive bands. These bands are mainly located in the region of near-infrared (NIR) and short-wave infrared (SWIR) [
8]. A combination of NIR and SWIR is necessary to retrieve canopy water content at leaf level [
9]. Previous studies have developed many vegetation indices, such as crop water stress index (CWSI) [
10], normalized difference water index (NDWI) [
11], ratio index [
12], etc. These indices are validated on the scale of satellite remote sensing, airborne remote sensing, and ground remote sensing. To reveal the relationship between land surface temperature and canopy water content, thermal and vegetation indices were integrated to invert the canopy water content, and the accuracy was higher than that used only one of them [
13]. The combination of thermal images and hyperspectral data can more precisely invert the canopy water content [
14]. For radiation transmission models, the PROSPECT model and PROSAIL model can retrieve canopy water content more accurately [
15]. The combination of hyperspectral data and PROCEST model has high inversion accuracy of canopy water content [
16]. However, when it was applied to large area of canopy water content inversion based on satellite remote sensing data, the accuracy will be low.
Synthetic-aperture radar (SAR) signals are strongly-penetrating and are unaffected by bad weather conditions, which makes SAR a useful tool for long-term agricultural monitoring in diverse environments [
17,
18]. SAR signals can penetrate crop canopies, potentially overcoming the phenomenon of optical data saturation, which can occur in high density vegetation. For moderately vegetated land surfaces, such as agricultural fields, the accuracy of the soil moisture estimation decreases due to the sensitivity of microwave signals to canopy water content [
19,
20]. These features, plus the sensitivity of scattered SAR microwaves to soil and vegetation characteristics, have led many researchers to explore SAR applications in agricultural monitoring. Many scholars carried out studies of soil water content inversion in cropland with vegetation cover. Water cloud model (WCM), as a semi-empirical model, was a classical vegetation contribution model [
21]. It treated vegetation as a continuous and evenly distributed aggregate, then the effects of vegetation have been addressed, and soil water content in vegetation-covered areas has high inversion accuracy. Fusion of SAR and optical data, vegetation index, and WCM were used to invert soil moisture in vegetation-covered areas, the inversion accuracy is higher than that of SAR alone [
22]. By the fusion of radiometer brightness temperatures and radar backscatter, it can reach a high soil moisture inversion accuracy, but on a lower spatial resolution [
23]. In order to invert soil moisture in crop growing areas, soil moisture index (SMI) was created, and it can better monitor changes in soil water content in the near-surface of agricultural areas [
24].
Traditional microwave remote sensing methods for monitoring the water content of vegetation are less. Empirical and semi-empirical models evolved earlier than physical models because they require fewer parameters. The inversion accuracy of vegetation water content based on empirical model is high, but it is limited to the study area [
25]. Physical models can retrieve vegetation water content more accurately and used minimum data of sample points. It can be used for any vegetation type and SAR data type [
26]. Recently, the study of vegetation water content monitoring based on passive microwave sensors has also been proposed. However, backscatter is greatest sensitivity to leaf moisture, but the trunk moisture is significant at low values of leaf moisture content [
27]. This shows that Vegetation Optical Depth (VOD) is sensitive to changes in vegetation water content, as plants respond to water stress [
28]. In agriculture and based on active microwave sensors, researchers investigated the effect of diurnal variation in vegetation water content of an agricultural canopy on backscatter for different radar configurations, the results showed the very significant effects that vegetation water content dynamics have on radar backscatter [
29]. Most of the above studies were based on passive microwave sensors for vegetation water content monitoring in a large scale. However, there were few studies on quantitative inversion of crop canopy water content based on active microwave sensors in a plot scale. Moreover, the complementary remote sensing data sources for monitoring crop water content have not been discussed.
Research on combining optics and SAR for remote sensing of vegetation has also been conducted [
30]. However, comparative analyses of optical and SAR-based vegetation water content models are lacking, especially those that use data from the Sentinel-1 satellite. Sentinel-1 is a dual-polarized radar remote sensing satellite with the potential ability to rapidly and continuously monitor crop water content [
31]. A model for monitoring winter wheat crop water content that is constructed from dual-polarized Sentinel-1 remote sensing data could enable early detection of water shortages in winter wheat crops, and ultimately provide decision support in agricultural production. Sentinel-2 is an optical satellite carrying a multi-spectral imager (MSI) with 13 bands as its main payload, and it can effectively monitor crop growth [
32,
33].
In this study, Sentinel-1 and Sentinel-2 images were used to construct and compare models for monitoring winter wheat crop water content. The combination and complementarity of the two data sources provide a feasible method for monitoring winter wheat crop water content due to the short satellite revisit period, rich data sets, and reliability in bad weather. The model based on Sentinel-1 and Sentinel-2 data enables accurate and rapid estimates of winter wheat crop water content over large land areas. The structure of this paper is as follows. In
Section 2 the study area, data sources, and the method used to verify model accuracy were described. In
Section 3 the results of model accuracy verification and affecting factors for model estimates were analyzed. In
Section 4 the findings were further discussed. In
Section 5 the main findings of this study were summarized, which provide a reference for future research.
3. Results
3.1. Choosing the Optical Vegetation Index and SAR Polarization Index
In this study, 28 optical vegetation indices were selected and 23 enhanced radar polarization indices were proposed.
Table 3 shows the results of the correlation analysis between the 28 optical vegetation indices and the water content of winter wheat at the sample points. The correlation between the optical vegetation indices and the measured water content of wheat ear and crop is basically similar. The correlation analysis results showed that the optical vegetation indices gerenally had more higher correlation with measured wheat crop water content than the measured wheat ear water content. The correlation between the optical vegetation indices and the measured water content of wheat stem and leaf is generally low.
The vegetation indices with the highest degree of correlation with the water content of winter wheat stem and leaf were SIWSI3 and MSI2 (Pearson correlation coefficients were 0.431 and −0.411, respectively). The vegetation indices with the highest correlation with the water content of winter wheat ears were NDWI, NDVI, and SIWSI3 (Pearson correlation coefficients were −0.513, 0.487 and 0.480, respectively). The vegetation indices with highest correlation with the water content of winter wheat crop were SIWSI3, MSI2 and NDVI2 (Pearson correlation coefficients were 0.607, −0.575 and 0.562, respectively). Correlation analysis revealed that the optical vegetation indices based on Sentinel-2 data generally correlated well with the water content of winter wheat crop.
The water content of winter wheat crop and the water content of winter wheat ear were closely same under different vegetation indices. In contrast, the vegetation indices did not correlate well with the water content of winter wheat stem and leaf. This is because optical remote sensing makes it easy to obtain crop canopy spectral reflectance information, but has limitations on the acquisition of internal information of crop (like stem). In the later stages of crop growth, vegetation biomass and canopy coverage are both high. Optical remote sensing cannot easily detect the water content of stems and leaves because of the poor permeability of the canopy. Nonetheless, optical remote sensing effectively detects the water content of winter wheat ears and canopy. Study observed that the red edge band, near-infrared band, and short-wave infrared band of the Sentinel-2 satellite were highly sensitive to the water content of the winter wheat crop, suggesting that the Sentinel-2 data are suitable for establishing vegetation indices for used to estimate crop water content.
The method (see
Section 2.4) proposed in this study was used to construct 23 enhanced radar polarization indices. In order to select the optimal radar polarization index, a correlation analysis between the indices and ear water content, stem and leaf water content, crop water content, and soil water content was carried out. The correlation between the enhanced radar polarization indices and soil water content was also analyzed (
Table 3).
Correlation between the enhanced radar polarization indices and soil water content was generally low when vegetation coverage was high. Only the correlations between the polarization indices (, and ) and soil water content reached significant levels (Pearson correlation coefficients were −0.355, −0.293 and −0.279, respectively). Most of the other radar polarization indices did not correlate well with soil water content. This is because radar signals produce complex responses to soil moisture in areas covered by vegetation. During the grain filling stage, vegetation coverage is high. This leads to complex scattering of the radar signal between vegetation layers, and therefore complex and changeable surface information contained in the scattered signal. The selected enhanced radar polarization index () significantly negatively correlated with soil water content in vegetation-covered areas.
3.2. Inversion of Winter Wheat Water Content Using Sentinel-2 Data
Table 4 shows the gray relational analysis for the five vegetation indices that correlated best with the water content of winter wheat. They are successively from first to last: NDWI, NDVI, SIWSI3, SMI2 and NDVI2. Used gray relational analysis, three of the indices (NDWI, NDVI and SIWSI3) were selected to build the water content inversion model, which is expressed as Formula (14).
Figure 3 presents the regression model for winter wheat water content built from the five vegetation indices, which shows that each model produced relatively similar precision.
Figure 4 shows the inversion results of winter wheat crop water content model which was constructed by multiple linear regression based on Sentinel-2 data. The model randomly selected 38 sample points and produced an accuracy of R = 0.644, RMSE = 0.018, nRMSE = 14.89%. The verification step using 20 verification sample points produced an accuracy of R = 0.632, RMSE = 0.021, nRMSE = 19.65%. The Sentinel-2-based model results were better than the inversion results using a single vegetation index.
3.3. Inversion of Winter Wheat Water Content Using Sentinel-1 Data
38 randomly selected sample points (the modeling set) were used to establish a water content inversion model for winter wheat using data from the Sentinel-1 satellite. The Levenberg-Marquardt and global optimization algorithm was used to estimate the fitting coefficients of the water cloud model. The fitting coefficients and model accuracy are shown in
Table 5. Model accuracy was R = 0.471, RMSE = 0.022 and nRMSE = 19.98%. 20 sample points were used to verify model accuracy and the results (R = 0.433, RMSE = 0.026 and nRMSE = 21.24%) were similar to those for model accuracy. The distribution of sample points in the verification set is shown in
Figure 5.
The verification step demonstrated that the model accurately estimated the water content of wheat crop. Model-estimated crop water content for wheat crop sample points located in the middle of the ground-measured crop water content interval. For both high and low values for crop water content, the model-estimated values for crop water content tended to be lower and higher than ground-measured values. This demonstrates that the inversion effect of the model for high or low values for crop water content is not particularly good. The reason for this phenomenon may be that the fitting accuracy of model coefficients is not good. Because the model set was randomly selected, some errors exist in the correction of model coefficients in the sample points, leading to large deviations in the verification results for extreme values in the verification set.
The radar polarization index, which is limited to the dual polarization of the Sentinel-1 image data, generally does not correlate well with soil water content in vegetation-covered areas. Furthermore, the overall inversion accuracy of the model is reduced. The model exhibits large error when inverting high and low values for crop water content. However, model error was within an acceptable error range. These results indicate that Sentinel-1 radar data have the potential to detect the water content of wheat crop.
3.4. Mapping the Water Content of Winter Wheat Crops Using Remote Sensing Data from Sentinel Series Satellites
In this study, two models for monitoring winter wheat crop water content based on data from the Sentinel-1 and Sentinel-2 satellites were proposed. The models based on the satellite data to estimate the water content of wheat crops in Gaocheng district. Thematic maps of wheat crop water content are presented in
Figure 6. Sentinel-1 imagery was available for the entire study site, whereas some Sentinel-2 imagery was missing for parts of the study area.
Gaocheng District’s wheat planting regions are mainly concentrated in the south and north. Winter wheat water content was highest in central Gaocheng District on 25 May 2017, followed by the north. The southern region exhibited the lowest wheat water content and was also the largest by land area.
The reason is that the wheat crops in central Gaocheng District receive plentiful irrigation due to the superior infrastructure that accompanies the region’s high population concentration and degree of urbanization. Remote monitoring estimates of wheat crop water content were largely consistent with traditional field survey results. The water content estimates based on SAR and optical remote sensing data were relatively consistent with each other, suggesting that Sentinel-1 imagery can be used for wheat crop water content monitoring in large areas when Sentinel-2 imagery is unavailable or impacted by weather conditions. The final remote sensing thematic maps have a good performance on water content of wheat crop.
4. Discussion
Correlation analysis between three enhanced radar polarization indices that are most sensitive to soil moisture and the water content of wheat ear, wheat stem and leaf, wheat crop, and soil in this study was showed in
Table 6. Enhanced radar polarization indices are less sensitive to crop canopy water content than soil water content. Because removed the influence of vegetation water content through the use of the water cloud model did not significantly affect the radar backscattering coefficient, and it showed that the backscattering coefficient directly extracted by Sentinel-1 radar images in vegetation-covered areas, such as farmland, can directly retrieve soil moisture and that vegetation had little effect on soil moisture inversion [
44]. The enhanced Sentinel-1 radar polarization indices correlated better with soil water content than did the original radar polarization indices. This suggests that the enhanced radar polarization indices constructed from dual-polarized Sentinel-1 radar image in our study can retrieve accurate estimates of soil moisture, even in regions covered by crops. Through compared two polarization parameters, VH and VV, of the Sentinel-1 radar images and found that the inversion of vegetation water content based on VV was more accurate than that based on VH. Previous study found that VV was more sensitive to soil moisture [
19]. It shows that the VV of dual-polarized radar data is more sensitive to vegetation water content and soil water content than VH when using water cloud model.
The result of wheat crop water content inversion based on Sentinel-1 showed that modeling sample points and verification sample points all showed substantially consistent model accuracy in our study. Compared with the previous study that simulated C-band polarization data, and then invert the vegetation canopy water content based on semi-empirical model constructed by cross-polarization ratio, incidence angle, and frequency [
25], our study has relatively higher inversion accuracy. This shows that the C-band enhanced radar polarization index combined with water cloud model has a well accuracy in the inversion of vegetation water content. Previous study has also found that the L-band RVI was highly correlated with wheat canopy water content [
45]. The inversion result of wheat canopy water content based on C-band radar is similar to that of the above study, which indicates that the proposed method in our study can effectively estimate crop wheat canopy water content in farmland areas. In addition, previous studies used the brightness temperature of passive microwave to invert vegetation water content, compared with our study, the higher accuracy of vegetation water content inversion can be obtained from these [
46,
47]. However, Sentinel-1, as an active microwave sensor, has higher spatial resolution compared with passive radar. It indicates that Sentinel-1 can be used to invert vegetation water content at field scale. In previous study, based on the Compact-polarimetric (CP) SAR data simulated by Radarsat-2 full-polarization SAR data, the rice canopy water content was inverted using CP backscattering coefficients and the water cloud model, which obtained a very high inversion accuracy [
48]. Compared with the inversion results of CP SAR, the inversion accuracy of Sentinel-1 data for crop canopy water content in our study is lower. It is because that the polarization index of Sentinel-1 data confined to dual-polarization is generally less correlated with soil moisture content than full-polarization SAR data in agricultural areas [
49,
50].
In our study, determination of the expression of model parameters was based on the principle of least fitting coefficients and did not attempt linear regression or nonlinear regression expression between sensitive SAR polarization indices and soil water content. In future work, a reasonable parameter expression form will be determined to obtain more accurate model inversion results.
5. Conclusions
In this paper, the results showed that the wheat crop water content can be effectively estimated based on water cloud model and enhanced radar polarization index by using Sentinel-1 dual-polarized radar data in the absence of optical data. However, the wheat crop water content inversion accuracy based on Sentinel-1 data is lower than that of based on Sentinel-2 data. This indicates that Sentinel-1 data can be used as a supplementary data source for the inversion of the wheat crop water content.
Here, two inversion models for winter wheat crop water content based on Sentinel-1 SAR and Sentinel-2 optical images were establised. 58 ground sample points were used for modeling and verification. For Sentinel-1 data, 23 enhanced radar polarization indices were firstly constructed and then selected most sensitive index to the measured soil water content as an input parameter to the water cloud model to retrieve the wheat crop water content. The verification accuracy of the inversion model based on Sentinel-1 data was R = 0.433, RMSE = 0.026 and nRMSE = 21.24%. For Sentinel-2 data, the gray relational analysis was used for several optical vegetation indices, and three vegetation indices were selected for regression modeling to retrieve the wheat crop water content. The verification accuracy of the inversion model based on Sentinel-2 data was R = 0.632, RMSE = 0.021 and nRMSE = 19.65%. The inversion accuracy of the Sentinel-2-based model was also higher than that of the Sentinel-1-based model. Because SAR can penetrate the vegetation canopy, the Sentinel-1 SAR-based method has the advantage of being able to estimate wheat crop water content during bad weather, such as cloud cover, which negatively affects optical data quality. Based on study results, the use of Sentinel-1 SAR data for the estimation of water content in wheat crops deserves further study. In addition, SAR data and optical data may be combined to explore more effective inversion methods for crop water content.