All-Sky Soil Moisture Estimation over Agriculture Areas from the Full Polarimetric SAR GF-3 Data

Abstract

This study aims to estimate the soil moisture (SM) in all-sky agriculture areas using fully polarimetric synthetic aperture radar (SAR) Gaofen-3 (GF-3) data. The radar vegetation index (RVI) is obtained using the radar SAR data, which overcomes the difficulty that the optical data cannot construct the vegetation index in cloud-covered areas. The RVI is introduced into the water cloud model (WCM) to remove the contribution of vegetation to the total radar backscatter ${\sigma }^0$ and obtain the soil backscattering coefficients ${\sigma }_{\mathrm{soil}}^0$ with HH and VV polarization. Subsequently, ${\sigma }_{\mathrm{soil}}^0$ and radar frequency data are introduced into the Chen model, and a semi-empirical model of SM estimation is established. The main findings of this study are as follows: (1) Compared with the ${\sigma }^0$, the ${\sigma }_{\mathrm{soil}}^0$ obtained by the WCM has a stronger correlation with the SM. (2) In the cloud covered area, the accuracy of the estimated SM by synergistically using the WCM and the Chen model is ideal. An RMSE of 0.05 and a correlation coefficient (r) of 0.69 are achieved. In this study, the SM estimation method is not affected by clouds, and it shows many advantages for sustainable development, monitoring soil drought degree, and other related research.

1. Introduction

Soil moisture (SM) is a critical factor in energy exchange between land and atmosphere and plays an essential role in the water cycle [1,2]. SM takes on a critical significance in ecology, hydrology, agriculture, and climate change research. Large-scale SM monitoring is beneficial in solving the problems relating to crop yield estimation, agricultural irrigation, drought monitoring, and seasonal climate evolution and prediction [3,4,5]. The conventional method of measuring SM is based on point measurement. Its measurement results exhibit adverse characteristics (e.g., temporal and spatial discontinuity and high cost), thus hindering large-scale SM estimation and dynamic monitoring [6,7].

Remote sensing (RS) technology is a novel method to monitor the SM in a large area. Microwave (MW) RS shows the advantages of complete coverage and all-weather observation [8,9,10]. Synthetic aperture radar (SAR), as an active MW sensor, can provide high temporal and spatial resolution monitoring data under any weather conditions [11,12]. The backscattering characteristics measured by the SAR are affected by specific factors (e.g., dielectric properties), and the soil dielectric constant is directly related to SM. Accordingly, the SAR can directly indicate the content and change of the SM [13,14,15].

The SM estimation models based on SAR data mainly include physically-based models, empirical models, and semi-empirical models [16]. In the bare land, the common models include the physical Integral Equation Model (IEM) [17], semi-empirical models of Dubois [18] and Oh [19] as well as their improvements (e.g., Advanced Integral Equation Model (AIEM) [20,21] and the Chen model [22]). For vegetated areas, the common models include the physical Michigan Microwave Canopy Scattering model (MIMICS) [23] and the semi-empirical Water Cloud Model (WCM) [24].

The WCM has been the most extensively used scattering model for estimating farmland SM from SAR data [25,26]. Backscatter analysis should consider the effect of soil, but also the effect of vegetation coverage and the interaction between the ground and plants [27,28]. The WCM should obtain the vegetation parameter input to characterize the vegetation water content [29,30]. In practical applications, optical RS data are usually used to obtain vegetation parameters [31]. The vegetation index obtained by Landsat-8 and Sentinel-2 data can be applied to the WCM to remove the contribution of vegetation to radar backscatter [32,33]. However, the optical data cannot apply to the cloud-covered areas. Moreover, optical and radar data are difficult to obtain from the same period. The radar vegetation index (RVI) extracted from SAR data is significantly correlated with leaf area index and vegetation water content, which can be adapted to describe vegetation parameters [34,35].

The WCM uses simulated or measured surface roughness and vegetation coverage datasets to simplify the theoretical backscattering models. However, surface roughness is highly difficult to measure when estimating the SM over wide areas [36]. Moreover, the main factors for the detection of the SM by radar include soil dielectric constant, surface roughness parameters, vegetation cover, etc. [37,38]. In vegetation-covered areas, the composition of radar signals is highly complex, and the SM is difficult to estimate.

The Chen model [22], modified from the IEM model [17], is suitable for large incidence angle ranges and wide roughness ranges. The Chen model assumes that the surface roughness can be expressed by an exponential correlation equation. In addition, this model considers the effect of radar incidence angle and frequency on backscattering. This study aims to make synergic use of the WCM and the Chen model to eliminate the effect of vegetation, surface roughness, radar incidence angle and radar frequency on radar backscatter signal, so this method can be applied to SM inversion in vegetated areas (e.g., farmland). Based on the SAR Gaofen-3 (GF-3) data, the RVI is obtained and brought into the WCM to obtain the soil backscattering coefficient. Subsequently, the soil backscattering coefficient, radar incidence angle and frequency data are introduced into the Chen model, and a semi-empirical model of SM estimation involving only radar parameters is built. The SM in the cloud-sheltered area is estimated based on the combination of the measured SM and SAR data.

2. Study Area and the Data Source

2.1. Study Area

The study area (113.67°–114.32° E, 32.69°–33.81° N) is on the southwest edge of the North China Plain (in Henan province, China); a few areas in the south are low hills, and the rest are plains (Figure 1). The study area is located in the transition zone between the subtropical and temperate zones. The interannual precipitation variation is considerable, with uneven temporal and spatial distribution. In general, there is less precipitation in winter and spring, and more precipitation in summer and autumn, mainly concentrated in July and August.

The soil types in the study area primarily include alflsols (soil with obvious clay leaching) and semi-aquatic soil (soil with temporary stagnant water). Crops (e.g., corn and peanuts) are mainly planted on the farmland in the region. The demand of crops for water is different in different periods. Timely monitoring of the degree of soil dryness positively affects agricultural irrigation.

2.2. Gaofen-3 (GF-3) Data

The GF-3 satellite is a C-band multi-polarimetric synthetic aperture radar satellite, which shows the characteristics of multi-imaging mode, large imaging width and high resolution [39]. The revisit period is 2–3 days, the radar incidence frequency is 5.4 GHz, the conventional incidence angle is 20–50°, and the extended incidence angle ranges from 10° to 60° [40]. The GF-3 offers full polarimetric data downloaded from the Earth Observation and Data Center, China National Space Administration (https://www.cheosgrid.org.cn/, accessed on 18 September 2020) for different product types and acquisition modes. The GF-3 SAR image employed is single look complex (SLC) data with the spatial resolution of 8 m × 8 m, which provided HH, HV, VH and VV polarization data. The GF-3 data were obtained on 7 August 2019.

2.3. In Situ Measurements

The in situ measurements data employed in the study are from the China National Meteorological Science Data Center (https://data.cma.cn/, accessed on 15 October 2020). The meteorological bureau uses the GStar-I (DZN2) automatic SM observation instrument to collect hourly soil relative moisture data [41]. The hourly SM curve of each station can test the integrity of the data of this station. Abnormal values can be eliminated through the data change trend to ensure the accuracy of the measured SM [42]. The SM with an underground depth of 10 cm–100 cm is examined. The in situ measurements data with a depth of 10 cm are used to build the model and verify the accuracy of the estimated SM.

The SM data acquisition time applied in this study is consistent with the imaging time of the GF-3 data. There are 20 SM observation stations in the study area. The minimum, mean, maximum, and standard deviation of the measured soil relative moisture values were 25.9, 87.1, 53.7, and 18.7%. Combined with the measured SM data, the spatial distribution map of SM with a spatial resolution of 8 m × 8 m is generated using the Kriging interpolation method (Figure 2). In the study area, the SM value tends to decrease from north to south, in a range from 23.4% to 96.4%.

3. Methods

The GF-3 image is preprocessed by radiometric calibration and terrain radiometric correction to obtain the radar backscattering coefficient (${\sigma }^0$) and radar incidence angle ($\theta$). The RVI is obtained using GF-3 data and brought into WCM to estimate the vegetation backscattering coefficient (${\sigma }_{{}^{\mathrm{veg}}}^0$). The soil backscattering coefficient (${\sigma }_{\mathrm{soil}}^0$) can be calculated by subtracting the ${\sigma }_{\mathrm{veg}}^0$ from the ${\sigma }^0$. The ${\sigma }_{\mathrm{soil}}^0$, $\theta$ and radar frequency data were put into the Chen model to remove the effect of soil roughness, radar incidence angle and frequency. A semi-empirical model for estimating SM is built by synergistic use of the WCM and the Chen model. Figure 3 presents the flow chart for estimating the SM.

3.1. Radar Vegetation Index (RVI)

Kim and Van [43] proposed combining the radar image intensity information to reflect the vegetation condition. In the study of estimating SM using radar data, the radar vegetation index based on the intensity value is defined (Formula (1)). The RVI can well indicate the vegetation information, and this study uses the RVI to estimate the vegetation water content. When the surface is smooth bare soil, the RVI is close to 0. With the growth of vegetation on the surface, the RVI tends to increase.

$$RVI=\frac{8{\sigma }_{\mathrm{HV}}^0}{2{\sigma }_{\mathrm{HV}}^0+{\sigma }_{\mathrm{HH}}^0+{\sigma }_{\mathrm{VV}}^0}$$

(1)

where ${\sigma }_{\mathrm{HV}}^0$, ${\sigma }_{\mathrm{HH}}^0$ and ${\sigma }_{\mathrm{VV}}^0$ are the HV, HH and VV polarization backscattering coefficients, respectively.

3.2. Water Cloud Model (WCM) and Chen Model

The vegetation layer will contribute to the ${\sigma }^0$ in the area covered by vegetation. The effect of vegetation on the backscattering coefficient must be subtracted when retrieving the SM from the SAR data. The WCM formula is as follows:

$${\sigma }_{}^0={\sigma }_{\mathrm{veg}}^0+{\gamma }^2\cdot {\sigma }_{\mathrm{soil}}^0$$

(2)

$${\sigma }_{\mathrm{veg}}^0=A\cdot m_v\cdot \cos \theta \cdot \left(1-{\gamma }^2\right)$$

(3)

$${\gamma }^2=\exp \left(-2B\cdot m_v\cdot \sec \theta \right)$$

(4)

where ${\gamma }^2$ is a double attenuation factor for radar waves passing through vegetation; m_v is the vegetation water content; and the values of A and B depend on the vegetation type.

Hence, the soil backscattering coefficient can be expressed as:

$${\sigma }_{\mathrm{soil}}^0=A\cdot m_v\cdot \cos \theta +\frac{{\sigma }^0-A\cdot m_v\cdot \cos \theta }{\exp (-2B\cdot m_v\cdot \sec \theta )}$$

(5)

Based on the IEM model, Chen assumes that the soil roughness can be expressed by an exponential correlation equation. The Chen model adopts the ratio of backscattering coefficients of HH and VV polarization to express the backscattering characteristics of the earth’s surface. The Chen model is expressed as:

$$\ln SM=C_1\cdot \frac{{\sigma }_{\mathrm{HHsoil}}^0}{{\sigma }_{\mathrm{VVsoil}}^0}+C_2\cdot \theta +C_3\cdot f+C_4$$

(6)

where ${\sigma }_{\mathrm{HHsoil}}^0$/${\sigma }_{\mathrm{VVsoil}}^0$ is the ${\sigma }_{\mathrm{soil}}^0$ of HH/VV polarization, f is radar frequency (GHZ), and C₁~C₄ are the coefficients which need to be fitted.

Finally, a semi-empirical model for the SM estimation was developed:

$$SM=\exp \left\{C_1\cdot \frac{\left[A\cdot m_v\cdot \cos \theta +\frac{{\sigma }_{\mathrm{HH}}^0-A\cdot m_v\cdot \cos \theta }{\exp (-2B\cdot m_v\cdot \sec \theta )}\right]}{\left[A\cdot m_v\cdot \cos \theta +\frac{{\sigma }_{\mathrm{VV}}^0-A\cdot m_v\cdot \cos \theta }{\exp (-2B\cdot m_v\cdot \sec \theta )}\right]}+C_2\cdot \theta +C_3\cdot f+C_4\right\}$$

(7)

3.3. Validation

To evaluate the accuracy of estimated SM, two indicators, namely correlation coefficient (Correlation, r) and Root Mean Squared Error (RMSE), were used. The two indicators were calculated as follows:

$$r=\frac{{\displaystyle {\sum }_{i=1}^N\left(x_i-\overline{x}\right)\left(y_i-\overline{y}\right)}}{\sqrt{{\displaystyle {\sum }_{i=1}^N{\left(x_i-\overline{x}\right)}^2\sqrt{{\displaystyle {\sum }_{i=1}^N{\left(y_i-\overline{y}\right)}^2}}}}}$$

(8)

$$RMSE=\sqrt{\frac{{\displaystyle {\sum }_{i=1}^N{(y_i-x_i)}^2}}{N}}$$

(9)

where x is estimated SM, y is measured SM, a and b are the average values of estimated and measured SM, N is the number of the sample. In general, the larger the r value, the smaller the RMSE, indicating the better the estimation effect.

Since this study is based on sparse coverage of the SAR dataset, we adopted an approach of substituting soil moisture comparison across space instead of time-series over a location to validate the soil moisture retrievals on a daily basis. In this approach, an ergodic substitution of space for time is adopted to match the grids of SM retrievals and in-situ SM measurement for a particular day with the assumption that those locations are geophysically similar in characteristics (i.e., biases in the grids being similar due to spatial autocorrelation) and mimic the time-series with different soil moisture states [44,45].

4. Results

4.1. Soil Backscattering Coefficient (${\sigma }_{\mathrm{soil}}^0$)

To compare the effect of vegetation factors on radar backscatter, Figure 3 shows the distribution of backscattering coefficient (${\sigma }^0$) and soil backscattering coefficient (${\sigma }_{\mathrm{soil}}^0$). The high-value area of the ${\sigma }^0$ is urban buildings, and the low-value area is water bodies and partial hilly regions. Figure 4 shows the ${\sigma }_{\mathrm{soil}}^0$ after removing the effect of the vegetation layer by WCM. After removing the contribution of vegetation to radar backscatter, the value of the ${\sigma }_{\mathrm{soil}}^0$ is lower than ${\sigma }^0$, and the ${\sigma }_{\mathrm{soil}}^0$ of HH/VV polarization is between −14~0 dB, while the ${\sigma }_{\mathrm{soil}}^0$ of HV/VH polarization ranges from −18 to 0 dB. Comparing the spatial distribution of SM and ${\sigma }_{\mathrm{soil}}^0$, their values tend to decrease along the northwest-southeast direction. The radar is sensitive to SM due to the soil dielectric constant, which is one of the key parameters in the radar backscattering coefficient. Hence, the spatial distribution of the SM and ${\sigma }_{\mathrm{soil}}^0$ shows high consistency.

The high(>−2 dB) and low(<−6 dB) backscattering coefficient values of HH/VV polarization are cross-distributed in the hilly areas in the south of the study area. During SAR data imaging, in the area with steep slopes, the hilly areas facing the radar signal are affected by perspective shrinkage and overlay, thus causing too strong backscattering. However, the terrain blocks the mountainous area with a back radar signal, and the backscattering in this area is extremely low.

Table 1. Correlation coefficient of ${\sigma }^0/{\sigma }_{\mathrm{soil}}^0$ and in situ SM data.

Backscattering Coefficient	HH	HV	VH	VV
Backscattering coefficient (${\sigma }^0$)	0.3860	0.2941	0.3703	0.5874
Soil backscattering coefficient (${\sigma }_{\mathrm{soil}}^0$)	0.4235	0.3089	0.4163	0.6084

lists the correlation coefficient (r) between the ${\sigma }^0$/${\sigma }_{\mathrm{soil}}^0$ and the SM. The r between ${\sigma }_{\mathrm{soil}}^0$ and SM is higher than that between ${\sigma }^0$ and the SM. After the vegetation backscattering coefficient ${\sigma }_{{}^{\mathrm{veg}}}^0$ is removed, the correlation between radar backscatter and SM is improved. The co-polarized backscatter is most significantly correlated with the SM; the r between VV/HH polarization backscatter and SM are 0.6084 and 0.4235, respectively. The correlation between cross-polarization backscatter and SM is less significant. The above result suggests that the co-polarized backscattering coefficient is more conducive to estimate SM.

4.2. Soil Moisture Estimation and Verification

Half of the measured spatial distribution data of SM (Figure 2) and GF-3 data in the study area are substituted into Formula (6), and the formula is fitted for estimating SM:

$$\ln SM=1.01314\cdot \frac{{\sigma }_{\mathrm{HHsoil}}^0}{{\sigma }_{\mathrm{VVsoil}}^0}+0.06628\cdot \theta +4.19987\mathrm{E}13\cdot f-2.26793\mathrm{E}14$$

(10)

The SM is estimated by Formula (10) in the study area, and the spatial distribution of estimated SM (8 m resolution) is illustrated in Figure 5. The remaining half of the measured spatial distribution data of SM is significantly correlated with the estimated SM, with r = 0.63 and with the RMSE of 0.09. The comparison of the spatial distribution of estimated and measured SM suggests that the high values of SM are located in the northwest, whereas the low values of SM are located in the southeast of the study area. When the estimated SM is 13.9~30%, the correlation between measured and estimated SM will be the worst; r = 0.34. r is 0.71 and 0.64 between the measured and estimated SM when the estimated soil relative moisture is 30~90% and 90~115.7%.

In Figure 5, the high value (>90%) and low value (<30%) of soil relative moisture alternate in the hilly areas in the south of the study area. The radar signal is affected by large topographic relief, thus resulting in an inaccurate estimation of soil relative moisture. Synergistically using the WCM and the Chen model to estimate SM is more suitable for flat terrain areas.

Figure 6 presents the scatter diagram of measured and estimated soil relative moisture on 7 August 2019 at 20 observation stations. Affected by buildings, the ${\sigma }^0$ corresponding to one observation station is abnormal. After removing this set of data, there are 19 available data sets. The r between the estimated and measured soil relative moisture is 0.68942. The soil relative moisture with ideal accuracy can be estimated using the SM estimation semi-model from SAR data.

4.3. Extension of SM Estimation Model

The imaging mode of Sentinel-1 image used in this study is the interferometric wide swath (IW) mode that provides VV and VH polarization data. The dual-pol Radar Vegetation Index ($RVI=4{\sigma }_{\mathrm{VH}}^0/({\sigma }_{\mathrm{VV}}^0+{\sigma }_{\mathrm{VH}}^0)$) proposed by Mandal [35] is adopted to characterize the water content of vegetation. The soil backscattering coefficient of VH and VV polarization (${\sigma }_{\mathrm{VHsoil}}^0$ and ${\sigma }_{\mathrm{VVsoil}}^0$) are obtained from the water cloud model. The depolarization ratio ${\sigma }_{\mathrm{VH}}^0/{\sigma }_{\mathrm{VV}}^0$ is found to be highly sensitive to surface roughness [46,47]. Hence, ${\sigma }_{\mathrm{HHsoil}}^0/{\sigma }_{\mathrm{VVsoil}}^0$ representing the surface roughness in the Chen model is replaced with ${\sigma }_{\mathrm{VHsoil}}^0/{\sigma }_{\mathrm{VVsoil}}^0$. Lastly, the formula for calculating SM by VH/VV polarization backscattering coefficient is written as follows:

$$\ln SM=C_1\cdot \frac{{\sigma }_{\mathrm{VHsoil}}^0}{{\sigma }_{\mathrm{VVsoil}}^0}+C_2\cdot \theta +C_3\cdot f+C_4$$

(11)

Figure 7 depicts the location of the new study area (second case). There are 193 soil moisture observation stations in this area. After the missing and abnormal data are removed, there are 152 available SM data. The coefficients C₁~C₄ in Equation (11) are obtained by using the best-fitting method based on the 76 measured SM and Sentinel-1 SAR data. The remaining 76 sample data are adopted to verify the accuracy of the SM estimation model. Figure 8 depicts the scatter diagram of measured and estimated soil relative moisture. The estimated soil relative moisture achieves a similar accuracy as that estimated by GF3 data, indicating a r of 0.66029 on 7 July and a r of 0.67445 on 12 August. The results verify the applicability of the SM estimation model in the field of agriculture.

5. Conclusions

Based on the GF-3 radar data, the WCM and the Chen model are synergistically used to remove the effects of vegetation, surface roughness, radar incidence angle and frequency on radar backscatter, and a semi-empirical model for estimating SM is constructed. Compared with the radar backscattering coefficient (${\sigma }^0$), the soil backscattering coefficient (${\sigma }_{\mathrm{soil}}^0$) obtained by the WCM has a stronger correlation with SM. Putting the measured SM spatial distribution data, ${\sigma }_{\mathrm{soil}}^0$ of HH/VV polarization, radar incidence angle and frequency data into the Chen model enabled fitting and obtaining the formula for calculating the SM. Lastly, the SM is estimated in the study area, and the correlation coefficient between the estimated and measured SM is 0.63.

It is worth mentioning that the proposed model was modified and estimated the SM of different study areas in other periods by using Sentinel-1 data.

We modified the proposed model and combined sentinel-1 SAR data to estimate SM in other similar study areas. The estimated SM has a similar accuracy as that estimated by GF3 data. It shows that the SM estimated model built in this paper can be applied to other similar regions. To estimate SM in agricultural areas using the proposed model, we only need to update the coefficients of these indices using the corresponding training sample data.

Although the model is capable of estimating SM in cloud covered areas, the accuracy of the estimated SM is not ideal. The spatial distribution of SM data used in the model is obtained using the Kriging interpolation method based on the SM data provided by the observation stations. However, the spatial distribution of SM data cannot completely express the real SM. This study only validates SM in space, whereas the validated SM in time-series cannot be ignored.

The method of combining optical and radar data to estimate SM will be subject to the problems of inconsistent imaging time and spatial resolution, whereas the method of estimating SM using SAR data completely avoids the above problems. However, the application of SAR data is still limited by terrain. Areas with large topographic relief will result in the generation of abnormal radar signals, thus affecting the estimation results of SM. In the future research on estimating SM from SAR data, the effective removal of the effect of terrain on radar signals will be highly challenging.