Combined Use of GF-3 and Landsat-8 Satellite Data for Soil Moisture Retrieval over Agricultural Areas Using Artificial Neural Network

Soil moisture is the basic condition required for crop growth and development. Gaofen-3 (GF-3) is the first C-band syntheticaperture radar (SAR) satellite of China, offering broad land and ocean imaging applications, including soil moisture monitoring. ,is study developed an approach to estimate soil moisture in agricultural areas from GF-3 data. An inversion technique based on an artificial neural network (ANN) is introduced. ,e neural network was trained and tested on a training sample dataset generated from the Advanced Integral Equation Model. Incidence angle and HH or VV polarization data were used as input variables of the ANN, with soil moisture content (SMC) and surface roughness as the output variables. ,e backscattering contribution from the vegetation was eliminated using the water cloud model (WCM). ,e acquired soil backscattering coefficients of GF-3 and in situ measurement data were used to validate the SMC estimation algorithm, which achieved satisfactory results (R � 0.736; RMSE� 0.042). ,ese results highlight the contribution of the combined use of the GF-3 synthetic-aperture radar and Landsat-8 images based on an ANN method for improving SMC estimates and supporting hydrological studies.


Introduction
Soil moisture content (SMC) is an important parameter in hydrological, biological, agricultural, and other processes [1,2].Lower SMC can cause an increase in the bare soil surface, thus aggravating sandstorms [3][4][5].Many technological advances now allow efficient acquisition of soil moisture data.On the ground, the International Soil Moisture Network (ISMN) provides a global network of soil moisture in situ observations [6]. is network measures soil moisture at specific locations; thus, the data are in the form of discrete values as opposed to a soil moisture spatial distribution, although they provide temporally continuous observations [7].Microwave synthetic-aperture radar (SAR) collects data over a large area with high spatial resolution and provides an effective technological means of monitoring and assessing soil moisture.
Radar remote sensing is sensitive to soil moisture, due to the differences in the dielectric constants of soil and water; thus, the dielectric constant is one of the most important factors in the radar backscattering coefficient [8].Several physical and statistical models have been developed to estimate soil moisture.e best-known physical model is the Advanced Integral Equation Model (AIEM), which simulates the radar backscattering coefficients from SAR and various soil parameters (radar wavelength, polarization, incidence angle, soil dielectric constant, and surface roughness) [9].Statistical models based on experimental measurements are also widely used in soil moisture estimations.For bare soils, the most popular statistical models are those developed by Oh et al., which include an inversion diagram based on the cross-polarized ratio or copolarized ratio [10][11][12].e Dubois model uses HH and VV polarization data from multipolarized radar observations to estimate SMC [13].
e above-mentioned models are commonly applied to bare soil and cannot be applied directly in vegetation cover areas, due to the multiple scattering effects of vegetation canopies [14].
e water cloud model (WCM), a semiempirical forward model, generally assumes that the vegetation canopy is a uniform layer of a cloud of water droplets and has been widely used to separate out the contribution of vegetation backscatter [15][16][17][18].To minimize the effects of vegetation, many researchers have attempted to utilize optical remote sensing to obtain additional vegetation information [19][20][21].Meanwhile, other studies have demonstrated improved SMC estimation accuracy when optical and SAR data are combined, compared to estimates solely from SAR data [22,23].In addition, good performance for soil moisture estimation has been achieved using only one radar channel (one incidence angle and one polarization) [24][25][26]; in one study, there was no significant improvement in soil moisture estimation using two polarizations (HH and HV, C-band) as opposed to one [27].
e estimation of soil moisture is usually a nonlinear, illposed, complex process [28], which makes it suitable for artificial neural network (ANN) application.ANN is a model-free estimator, as it does not rely on an assumed form of the underlying data [29].e most direct way to train an ANN is using synthetic data generated by theoretical or empirical surface scattering models.e effectiveness of ANN inversion algorithms has been investigated in previous studies [30][31][32].Baghdadi et al. [33] tested the performance of an ANN in retrieving soil moisture and surface roughness for several inversion cases, with and without a priori knowledge of soil parameters; the results were promising for ANN soil parameter estimation.El Hajj et al. [34] developed and validated neural networks using synthetic and real databases; the application of both VV and VH showed results similar to those using VH only.Satalino et al. [35] combined an Integral Equation Model (IEM) and neural networks to retrieve SMC over smooth bare soils from ERS-SAR data.Paloscia et al. [36] trained neural networks using a backscattering coefficients database simulated from the IEM and WCM for a wide range of soil parameters; a real database composed of SAR, optical, and in situ measurements was used to validate the developed neural networks, and the results indicated a soil moisture estimation accuracy of 2-5 vol.%.Santi et al. [37] developed ANN-based algorithms for both active and passive microwave acquisitions; the results demonstrated that ANNs are a powerful tool for SMC estimation at both local and global scales.ese studies demonstrate the potential of ANNs for retrieving SMC information from SAR remote sensing data.
Gaofen-3 (GF-3) is the first Chinese civil C-band SAR.To date, there have been few soil moisture estimation methods developed for the GF-3 satellite.erefore, we create an inversion technique based on ANN to estimate SMC over agricultural areas by combining GF-3 and Landsat-8 satellite data.e WCM was first applied to eliminate the effects of vegetation and to obtain the backscattering coefficients of bare soil.
en, the ANN was trained using a sample dataset generated form the AIEM.Meanwhile, field SMC data in an agricultural region with wheat as the main crop type were used to evaluate the potential of the GF-3 sensor for retrieving SMC data.e remainder of this paper is organized as follows: Section 2 summarizes the study area and datasets.Section 3 discusses the ANN and the inversion methodology.e results are presented in Section 4. Section 5 includes a discussion of our findings and conclusions.1) was chosen to validate the approach for soil moisture estimation.e site is relatively flat, and the main crop type is wheat.e study area has a typical subhumid, north temperate continental monsoon climate.e annual average temperature is approximately 12.8 °C, and the annual average precipitation is approximately 474 mm, which is mainly concentrated in July and August.e top soil type of the agriculture fields is cinnamon soil with a high nutrient content, suitable for crop growth.Although the case study area is not large, it has the representative characteristics of crop-type distribution in the North China Plain.

Case Study Site and Data Description
GF-3 SAR and Landsat-8 OLI images were used in this study.In addition, 38 agriculture fields were selected to conduct in situ measurements, including those related to both soil and vegetation characteristics.

Data Description.
Microwave and optical satellite data were fused to reduce the effects of vegetation cover on the backscattering of soil moisture.Satellite data are listed in Table 1.

GF-3 Data.
In this study, one GF-3 SAR product acquired in the Quad-Polarization Strip I (QPSI) mode, processed up to level-1A SLC, was collected on May 27, 2017.Details of the QPSI model are listed in Table 2. e locations of the collected GF-3 SAR data are shown in Figure 1.PolSARpro software was used to calibrate the GF-3 image; the calibration aims to convert the digital number values of the GF-3 image into backscattering coefficients (σ °) in a linear unit.In this study, we focused only on the copolarizations of HH and VV.To reduce the effect of speckle noise, the mean backscattering coefficient of each sampling point was calculated from a calibrated GF-3 image by averaging the σ °values of five surrounding pixels.Meanwhile, to reduce the impact of residential area on the soil moisture mapping, a filtering method was used to mask houses (red patches in Figure 1) to simplify the soil moisture map.

Landsat-8 Image.
e NASA's Landsat-8 satellite carries two instruments: the Optical Land Imager (OLI) sensor and the ermal Infrared Sensor (TIRS).It images the 2 Advances in Meteorology land surface using 11 spectral bands in the optical and thermal infrared domains with a spatial resolution of 30 to 100 m and a temporal resolution of 16 days [38].e Landsat-8 imagery used in this study was downloaded from the United States Geological Survey data archive (https://earthexplorer.usgs.gov/).We directly downloaded the land surface re ectance product, in which the image had been preprocessed; preprocessing included radiation calibration and atmospheric correction.e re ectance values of near-infrared (NIR) and short-wave infrared (SWIR) bands were used to estimate the normalized di erence water index (NDWI) and VWC.Finally, Landsat-8 re ectance data were extracted from the sample points and combined with eld measurements to build the relationship between vegetation water content (VWC) and NDWI.

In Situ Measurement Data.
Coincident with the GF-3 and Landsat-8 satellite overpasses, eld campaign measurements of soil moisture and roughness, as well as crop biophysical parameters, were conducted over the 38 wheat elds.For each eld, three sampling points were randomly selected (point separation 150 m).Soil volumetric moisture was measured using the oven-drying method (wet weightdry weight) at a depth of 0-5 cm, given that the C-band radar signal is most sensitive to surface soil moisture.For each sampling point, measurements were collected at three locations that were uniformly distributed over 8 m (one pixel of the GF-3 satellite).e average soil moisture value of the three locations was considered the soil moisture of the sampling point.e VWC collected from 1.0 × 1.0 m squares selected at random was determined by weighting before and after oven-drying.Soil roughness was measured with a 1 m

Methodology
Our approach for the soil moisture estimation uses an ANN technique that combines GF-3 and Landsat-8 satellite data (Figure 2).e ANN was trained and tested on a training sample dataset generated from the AIEM.First, the WCM was used to eliminate the contribution of backscattering coefficients caused by vegetation.en, the backscattering coefficient of the soil was determined.e obtained soil backscattering coefficients of GF-3 and in situ measurement data were used to validate the SMC estimation algorithm.Finally, SMC was estimated using the trained ANN.

Calculation of Backscattering Coefficient of Soil
3.1.1.Vegetation Water Content Calculation.VWC (kg/m 2 ) is one of the most important parameters for the successful retrieval of SMC from microwave remote sensing observations [39].Landsat-8 Operational Land Imager (OLI) data and ground-based VWC measurements were used to establish relationships in our study based on remotely sensed indices.
e NDWI is a widely used and reliable indicator to assess the vegetation water status, which is sensitive to changes in VWC [40].Gao first proposed the NDWI by combining reflectance at 860 and 1240 nm to monitor VWC [41].Because large-area VWC is more difficult to obtain, NDWI was used in the current study.NDWI can be calculated as follows: where R SWIR is the reflectance or radiance corresponding to the SWIR wavelength channel (1.2-2.5 µm) [42].For Landsat OLI, R NIR and R SWIR correspond to bands 5 (0.845-0.885 µm) and 6 (1.560-1.660µm), respectively.VWC was measured in 38 fields at the first sampling point.e above ground biomass was removed, and fresh and dry weights were used to compute the VWC.e relationship between VWC and NDWI was generated based on the least-squares fitting method, as follows: where a and b are the coefficients and c is a constant calculated based on Landsat-8 OLI land surface reflectance data and ground-based VWC measurements.

Water Cloud Model.
e WCM, introduced by Attema and Ulaby [15], assumes that vegetation is a source of homogeneous scattering.Radar backscattering coefficient σ °from a canopy can be expressed as the sum of contributions due to (i) volume scattering σ °canopy from the vegetation canopy itself, (ii) surface scattering σ °soil by the soil attenuated by the vegetation layer, and (iii) multiple interactions σ °canopy+soil between the canopy and the ground surface [14].For a given incidence angle θ, the WCM can be represented as follows: where τ 2 is the two-way vegetation transmissivity.e interactions between vegetation and soil are neglected in the WCM [18,43]; therefore, the WCM can be reformulated as follows: where the backscattering coefficient of bare soil is simulated based on the AIEM, which will be introduced in Section 3.2.m veg is the VWC (kg/m 2 ).e incidence angle θ of GF-3 data used in this study was 24 °.A and B are parameters that depend on the canopy type and sensor configuration, which can be calculated by the least-squares method.

Generating the SMC Training Sample Dataset.
Bare soil backscattering depends on the dielectric constant and surface roughness, as well as the SAR instrumental parameters [44,45].e AIEM, a well-established theoretical model [9], has been widely used as a forward model to simulate the scattering coefficients and emissivity of bare soil surfaces with various ground conditions, due to its precision [44][45][46][47][48]. erefore, AIEM was selected to generate the SMC training sample dataset.
e Dobson dielectric model is commonly used to describe the relationship between the effective dielectric constant of soil and soil moisture [49][50][51].erefore, we combined the Dobson model and AIEM to integrate soil moisture into the training sample dataset during the generation process.e equation can be conceptually represented as follows: where f represents the satellite frequency (5.4 GHz for the GF-3 satellite), θ is the angle of incidence, PP denotes the polarization state (includes HH and VV polarizations), ACF is the autocorrelation function (an exponential ACF is adopted), s is the root mean square height, and l is the correlation length.
e incidence angle ranged from 20 °to 60 °with an interval of 1 °. e s and l values were set based on field measurements within 0.5-2.0cm and 10.0-30.0cm, respectively.To reduce the number of parameters, surface roughness is expressed as one parameter, Zs (s 2 /l), using an exponential correlation function.Soil moistures ranged 4 Advances in Meteorology from 0.01 to 0.40 m 3 /m 3 , with an interval of 0.01 m 3 /m 3 .e training sample dataset was generated based on the AIEM that included 551040 datasets.

Arti cial Neural Networks Approach.
ANNs can mimic human learning and can build multivariate nonlinear relationships; as such, they have been widely used for estimating land surface parameters from remote sensing data [52].An ANN is made up of a number of hidden neurons or nodes that work in parallel to convert data from an input layer into an output layer.Each ANN has two modes of operation: training and testing modes.In the training mode, neurons are trained using part of the training sample dataset as a particular input pattern to produce the desired output pattern.In the testing mode, when an input pattern is chosen, the ANN will produce its associated output [53].e number of neurons associated with the hidden layer varies, depending on the optimum neural network architecture.
Training is accomplished to obtain a minimum error between the ANN output and the input data by adjusting the correlation weights between them [54].e ANN model was developed using the MATLAB software.e incidence angle and backscattering coe cient (HH or VV) were the input variables; the corresponding SMC and surface roughness were the output variables.e ANN was trained for HH and VV polarizations separately.One hidden layer and 30 neurons provided accurate SMC estimation within a reasonable computing time, by adding or removing these components from the model for both HH and VV polarizations.
erefore, the optimal ANN architecture (Figure 3) was determined to be a three-layer network consisting of an input layer (two neurons: incidence angle and backscattering (HH or VV), one hidden layer (30 neurons), Advances in Meteorology and a two-output layer (SMC and surface roughness).Although the structures for both HH and VV polarizations are the same, the detailed ANN di ers between the two.HH and VV backscattering data were used separately as input parameters for their corresponding ANN. e optimum architecture has minimum error and maximum convergence, avoiding any possible over tting.e training sample dataset generated from the AIEM was randomly divided into two parts: 90% of the cases were used for training the ANN and the remaining 10% of the cases were utilized during the testing process.e Levenberg-Marquardt method, an alternative to the Newton algorithm, was used to calibrate the synaptic coe cients.Linear and tangent-sigmoid transfer functions were associated with the hidden layer and output nodes, respectively.

Soil Moisture Estimation and Accuracy Assessment.
GF-3 satellite data were preprocessed to obtain backscattering coe cients of the agricultural area.e backscattering coe cients of soil were generated based on the WCM to eliminate the backscattering contribution of vegetation.e SMC can be estimated using the trained ANN and the backscattering coe cients of soil as the input parameter.Direct comparison of in situ SMC measurements with SMC estimations using GF-3 satellite data is a reliable way to assess the accuracy of the proposed SMC estimation algorithm.e precision and accuracy of SMC were estimated using two statistical indices: the R 2 value of linear regression and the root mean square error (RMSE).e RMSE values show how much the retrieval SMC values under-or overestimate the in situ measurements.For a perfect t between retrieval and eld-observed SMC data, values of R 2 and RMSE should equal 1.0 and 0.0, respectively.

Vegetation Water Content.
Information about VWC is an important parameter of the WCM, which is useful for retrieving soil moisture from GF-3 satellite data.NDWI was chosen to generate relationships with VWC based on Landsat-8 OLI land surface re ectance data and ground-based VWC measurements.en, coe cients (a and b) and constant (c) of (2), 1.56, 1.27, and 0.49, respectively, were calculated using the least-squares tting method (R 2 0.771).e VWC estimation results using the proposed algorithm and Landsat-8 data are shown in Figure 4.As for the spatial distribution, the large VWC estimates were mainly distributed over the farmland, and the VWC estimates at other areas were smaller.Due to the di erent farmland areas, the growth of wheat is not the same, so there are di erences in VWC.erefore, the VWC estimates using Landsat-8 data could preliminarily indicate the reasonability of the proposed VWC estimation algorithm.

Correlations between In Situ SMC and Corresponding
Total Backscatter σ °.A sensitivity analysis between the GF-3 total radar backscatter σ °(HH and VV polarizations) and in situ SMC was conducted based on all the eld measurements data to explore whether SMC could be retrieved directly using regression methods, as shown in Figure 5. GF-3 total radar backscatter σ °(HH and VV polarizations) was correlated with SMC, which is consistent with previous ndings [55], demonstrating the potential of GF-3 satellite data for SMC retrieval.However, R 2 between σ °and in situ SMC, both with HH and VV polarizations, was lower than 0.146 (Table 3), thus indicating that simple regression methods cannot achieve high-precision inversion of SMC. e SMC estimation under the vegetation cover area is a ected by the vegetation canopies, which scatter and attenuate electromagnetic radiation, which makes it di cult to discriminate the radar return due to soil moisture [56].erefore, isolating the contribution of vegetation from the total radar backscatter is crucial for SMC estimation over agricultural areas.

Bare Soil Backscattering Coe cients σ °soil .
e bare soil backscattering coe cient σ °soil , which assumes vegetation as a homogeneous scattering source, was calculated using the WCM.
e simulated bare soil backscattering coe cient σ °soil − simu was developed based on AIEM.
en, the parameters A and B of the water cloud model were calculated through the least-squares method (Table 4).
e backscattering coe cients of 19 wheat elds (57 points) were extracted to analyze changes in σ °and σ °soil information (Figure 6).e backscattering coe cient value of each point was attenuated after the use of the WCM.However, the degree of attenuation of each point was not the same, mainly because the corresponding VWC was di erent.

AIEM-Simulated Backscattering Data σ °soil − simu .
SMC estimation was achieved using an ANN method, in which the training sample dataset was generated based on the AIEM.Using the trained ANN, GF-3 satellite-measured soil backscatter data were used as the input and SMC as the output.To verify the reliability of the training sample dataset, the consistency between the simulated backscattering data (σ °soil − simu ) and radar-measured soil 6 Advances in Meteorology backscattering data (σ °soil ) was explored based on the remaining 57 points, which is the di erent dataset used for calibrating the WCM, as shown in Figure 7. e results showed that both HH and VV polarization-simulated backscattering data from the AIEM agreed well with GF-3 satellite-measured soil backscattering data calculated using the WCM (R 2 0.894 for HH and R 2 0.855 for VV).
erefore, the training sample dataset generated from AIEM was applied to the SMC estimation in this study.

Soil Moisture.
e soil moisture estimation results are shown in Figure 8. e remaining 19 wheat elds, which contained 57 points, were used to directly assess the soil moisture from GF-3 data using the proposed algorithm (Figure 9).
ere was a good linear relationship between eld-measured soil moisture and estimated soil moisture.
e soil moisture retrieval accuracy was satisfactory, with R 2 and RMSE values of 0.7356 and 0.042 for HH polarization and 0.7096 and 0.051 for VV polarization, respectively.e main reason for some of the larger di erences between the in situ SMC and the estimated SMC may be attributable to the eld sampling period being out of sync with the satellite transit time, as the soil moisture could change over this time period.
In summary, the di erences between the measured soil moisture and estimated soil moisture from GF-3 data were small, and the soil moisture retrieval results were satisfactory.ese results indicate that the proposed soil moisture method for GF-3 data is reliable and that GF-3 data could achieve acceptable performance for soil moisture estimation.
us, this approach shows the potential for providing the high-resolution soil moisture dataset for agricultural application, such as farmland soil moisture monitoring.

Discussion and Conclusions
We propose a soil moisture retrieval algorithm for agricultural regions that uses GF-3 satellite and Landsat-8 data based on the ANN method.e ANN structure was trained under a large range of land surface parameters, allowing the algorithm to have better adaptability to a variety of underlying conditions.e retrieval results using eld soil moisture measurements obtained from an agricultural region with wheat as the dominated crop type showed that the proposed algorithm achieved satisfactory soil moisture estimation accuracy (e.g., RMSE 0.042).
e results indicated GF-3 satellite data had good performance on soil moisture retrieval, and the algorithm had potential to operationally estimate soil moisture from GF-3 satellite data.e major conclusions of this study are summarized as follows: (1) VWC is an important factor for accurate retrieval of soil moisture under the vegetation cover.In this study, we combined microwave and optical remote sensing data to eliminate the contribution of backscattering coe cients caused by the vegetation.were calculated using a least-squares method.(3) e backscattering coefficients of GF-3 satellite data were attenuated to different degrees compared to the total backscattering coefficients after the use of the WCM.Due to the different VWCs of each point, the attenuation degree is also different.e effect of vegetation contribution is large and must be removed before the soil moisture retrieval process; otherwise, it will influence the accuracy of soil moisture retrieval.(4) e optimal ANN architecture in this study was determined as a three-layer network consisting of an input layer (three neurons: incidence angle and HH or VV backscattering), one hidden layer (30 neurons), and a two-component output layer (SMC and surface roughness).Our results and model sensitivity highlight the contribution of combined GF-3 SAR and Landsat-8 images using an ANN method for improving SMC estimates.HH polarization showed better SMC estimation performance than VV polarization.
Although satisfactory soil moisture retrieval performance was achieved, there were also several limitations to this study.e field measurement had many uncertainties that affected assessing soil moisture retrieval algorithms using remote sensing data.Real soil moisture values from ground-based measurements are difficult to match with pixel-level soil moisture estimates, although averaging multipoint measurements could reduce this error to a certain extent.Further work should focus on validating the   Advances in Meteorology proposed soil moisture retrieval algorithm using eld soil moisture measurements with less uncertainty under various land cover-type conditions.
Data Availability e data used to support the ndings of this study are available from the corresponding author upon request.

Conflicts of Interest
e authors declare no con icts of interest.

Authors' Contributions
Linlin Zhang, Qingyan Meng, and Qiuxia Xie conceived and designed the experiments.Linlin Zhang and Shun Yao

Advances in Meteorology
performed the experiments and wrote the paper.Qingyan Meng, Xu Chen, and Ying Zhang contributed to paper revisions.

Figure 1 :
Figure 1: Left image shows the geolocation of the Luancheng study area in Shijiazhuang city, and the right image is Gaofen-3 (GF-3) Quad-Polarization Strip I (QPSI) data acquired on 27 May 2017.e red patches are residential areas, and the green triangles indicate the eld measurement sampling points (point separation: 150 m) in each eld.

Figure 2 :
Figure 2: Flow chart for soil moisture content (SMC) estimation using an arti cial neural network (ANN).

Figure 3 :
Figure 3: ANN architecture used for SMC estimation.

Figure 6 :
Figure 6: Changes in the backscattering coefficient (a) before and (b) after the elimination of vegetation information.

Figure 9 :Figure 8 :
Figure 9: (a, b) Scatter diagram between in situ soil moisture and retrieved soil moisture.

Table 1 :
Satellite data characteristics used in this study.

Table 2 :
Main technical speci cations of each imaging mode.pinplate, including the root mean square height (S) and correlation length (L) of each sampling point.In the 38 wheat fields, 19 fields (57 points) were chosen to complete the VWC calculation and validate the WCM, and the remaining 19 wheat fields (57 points) were used to validate the soil moisture estimation model using GF-3 data.

Table 3 :
Linear tting equations for in situ SMC and GF-3 radar backscattering coe cient σ

Table 4 :
Vegetation parameters in the water cloud model.
°soil .As the model's input parameters, the vegetation parameters A and B