Retrieving Soil Water Content in Winter Wheat Fields Using UAV-Based Multi-Source Remote Sensing and Machine Learning

Abstract

Retrieving farmland soil water content with both high accuracy and physical interpretability remains a significant challenge, particularly for winter wheat. To bridge the gap between purely empirical data-driven approaches and mechanistic scattering models, this study proposed a novel hybrid framework that integrates an improved water cloud model (IWCM) with machine learning algorithms. Multi-modal unmanned aerial vehicle (UAV) experiments were conducted during the heading stage of winter wheat over two consecutive years (2024–2025) using a synchronized system equipped with a miniature synthetic aperture radar (MiniSAR) and a multi-spectral sensor. The core innovation of the proposed framework lies in the IWCM, which explicitly decouples vegetation and soil scattering contributions by incorporating fractional vegetation cover, thereby deriving physically meaningful soil backscatter coefficients from complex microwave signals. Unlike traditional methods that treat remote sensing variables as black box inputs, our approach employed these physics-derived features to guide data-driven modeling. Four feature input schemes including spectral reflectance, vegetation indices, MiniSAR polarimetric parameters, and their multi-source fusion were systematically evaluated using back propagation neural network (BPNN) and random forest (RF) regressors. The results demonstrated that the proposed framework significantly enhances retrieval performance. Notably, the RF model driven by spectral band reflectance within this physically constrained architecture achieved optimal accuracy, with a coefficient of determination (R²) of 0.865, a mean absolute error (MAE) of 0.0152, and a root mean square error (RMSE) of 0.0197. Compared to purely empirical approaches, the IWCM significantly improved the physical interpretability of microwave polarimetric characteristics, enabling the multi-source data fusion to better represent the interactions among vegetation, soil, and microwave scattering. This study demonstrated that integrating mechanistic models with multi-source UAV remote sensing data not only improves soil water content retrieval accuracy in winter wheat fields but also provides a valuable reference for developing operationally applicable and physically interpretable farmland soil water content monitoring systems.

1. Introduction

Soil water content, as a critical parameter in agricultural production, plays a significant role in crop growth monitoring, precision irrigation and agricultural disaster early warning [1]. Traditional methods for monitoring soil water content primarily rely on manual field sampling and fixed-point instrument observations [2,3]. While these methods can provide accurate data, they are time-consuming, labor-intensive, and lack sufficient spatial representativeness at the regional scale, making them inadequate for field-scale dynamic management at the field level [4]. With the continuous advancement of remote sensing techniques, effective alternatives have emerged to overcome these limitations [5]. Among these, satellite remote sensing, with its advantage of large area coverage, has become the mainstream method for estimating soil water content at the regional scale [6]. Microwave remote sensing, capable of penetrating clouds and vegetation canopies, utilizes the physical relationship between backscatter coefficients and soil dielectric constant to retrieve soil water content through models such as the Oh model, IEM model, and machine learning algorithms [7,8]. Optical remote sensing (e.g., Landsat 8, Sentinel-2) indirectly corrects for vegetation interference on spectral signals by establishing statistical relationships between vegetation indices and soil water content [9]. Some researchers have further explored multi-source satellite collaborative inversion strategies [10]. For instance, Wang [11] fused microwave with optical remote sensing observations, while Prakash Mohan [12] employed the water cloud model (WCM) alongside an RBF neural network, using Sentinel-1/2 and Landsat 8 imagery as inputs. Both studies improved the estimation accuracy of surface soil water content to varying degrees.

However, satellite remote sensing still faces significant limitations in the fine-scale management of farmland [13,14]. Microwave satellites have a spatial resolution of approximately 10–20 m, which is insufficient to capture the spatial heterogeneity of soil water content caused by irrigation differences and micro-topographic variations within fields [15]. Optical satellites, on the other hand, are susceptible to atmospheric conditions and vegetation canopy interference [16], particularly after the crop jointing stage, where leaf occlusion leads to spectral saturation and a significant decline in inversion accuracy [17,18]. Additionally, the satellite revisit cycle (e.g., 12 days for Sentinel-1) creates a spatiotemporal mismatch with the high-frequency decision-making demands of agricultural production, such as irrigation scheduling and pest and disease early warning.

Against this backdrop, unmanned aerial vehicle (UAV) remote sensing has emerged as a key solution to the aforementioned challenges due to its advantages of high resolution, flexibility, real-time capability, and operational efficiency [19]. Equipped with multi-spectral [20,21], thermal infrared, and miniaturized synthetic aperture radar (MiniSAR) sensors, UAVs can achieve centimeter-level spatial resolution imaging [22,23]. Unaffected by cloud or fog obstruction, they accurately capture microscale characteristics of water distribution within field plots [24]. In recent years, significant progress has been made by researchers worldwide in integrating UAV remote sensing with machine learning. Shokati [25] achieved high-precision inversion of soil water content under different vegetation types by combining UAV-based hyperspectral remote sensing data with a random forest (RF) algorithm. Wang et al. [26] validated the feasibility of estimating soil water content (SWC) using UAV-based hyperspectral data and optimal models based on partial least squares regression (PLSR) and artificial neural networks (ANN). Zhang et al. [27] employed a portable L-band radiometer mounted on a UAV within a Bayesian inference framework to evaluate soil water content inversion across various land cover types, including shrubland, bare soil, and forest belts, demonstrating its potential for soil water content monitoring. Yin et al. [28] provided a reliable method for soil water content inversion in alfalfa fields using a deep neural network (DNN) model integrated with UAV multi-spectral data.

Despite the progress achieved in existing studies, several critical challenges persist in the field of remote sensing-based soil water content retrieval. A primary limitation is that most retrieval models continue to rely on single-sensor data and have yet to fully leverage the synergistic potential of multi-source remote sensing. Systematic reviews indicated that while passive microwave sensors offer high temporal resolution, their utility is constrained by coarse spatial resolution. Conversely, active microwave sensors provide finer spatial details but are limited by longer revisit cycles and sensitivity to surface roughness, necessitating advanced fusion strategies to harness their complementary advantages [29]. Singh et al. [30] demonstrated that multi-sensor synergy integrating Sentinel-1 radar backscatter, Sentinel-2 optical bands, and digital elevation model (DEM) topographic data can substantially enhance retrieval accuracy. Nevertheless, optical sensors, despite delivering high spatiotemporal resolution vegetation indices and land surface temperature data, exhibit limited penetration depth and are vulnerable to cloud contamination [31]. Microwave sensors, though capable of all-weather observation, are strongly influenced by canopy scattering in vegetated areas, making it challenging for any single data source to comprehensively capture the complex interactions within the soil–vegetation–atmosphere system. Moreover, prevailing multi-source data fusion approaches often remain superficial, largely limited to feature-level concatenation without deeper integration into physically consistent microwave scattering mechanisms. For instance, empirical parameters in classical models such as the WCM show significant variability across different crop types and phenological stages, introducing substantial uncertainty in parameterization [32].

Furthermore, although current machine learning models demonstrate strong capabilities in nonlinear fitting, their inherent “black-box” nature due to a lack of integration with physical mechanisms remains a significant limitation. Multiple studies have indicated that purely data-driven machine learning approaches may slightly outperform traditional physical or semi-empirical models in terms of predictive accuracy; hybrid methods that couple physical models with machine learning tend to achieve the highest accuracy [33,34]. However, purely data-driven models that lack physical constraints are prone to overfitting and exhibit limited generalization ability, particularly when training samples are scarce or under changing land surface conditions [35]. Moreover, persistent cloud cover affecting optical data further exacerbates the challenge of insufficient training samples.

In light of these considerations, this study focused on the winter wheat planting area as the research region in Xun County of Henan Province, China. Over two consecutive years (2024 and 2025), multi-temporal coordinated UAV observation experiments were conducted during the heading stage within the same study area. The research integrated MiniSAR microwave data with multi-spectral optical data for soil water content retrieval. Microwave remote sensing (e.g., Sentinel-1 SAR) has clear physical interpretability. The synergistic use of microwave and multi-spectral optical data effectively mitigated the limitations associated with single-sensor approaches. At the physical mechanism level, vegetation cover was introduced to refine the classical WCM, resulting in an IWCM that explicitly separates vegetation and soil scattering contributions. This allowed for the extraction of polarimetric soil backscatter coefficients. Furthermore, this study deeply integrated the IWCM with two machine learning algorithms—back propagation neural network (BPNN) and RF—by designing four distinct feature input strategies for comparative analysis: band reflectance, vegetation indices, MiniSAR polarimetric parameters, and multi-source composite variables. The objective was to develop a hybrid soil water content retrieval model for winter wheat that leveraged both physical constraints and data-driven advantages. Additionally, the flexibility and mobility of the UAV platform enabled on-demand observations, effectively alleviating the challenges associated with spatiotemporal matching in satellite data. The outcomes of this research are expected to provide technical support for winter wheat growth monitoring and precision irrigation, thereby contributing to intelligent agricultural management and enhancing food security.

2. Materials and Methods

2.1. Study Area

This study was conducted in Baisi Town, Xun County, Henan Province, China, within a well-facilitated farmland demonstration area. The region belongs to the warm temperate semi-humid monsoon zone, characterized by abundant sunlight, annual precipitation of approximately 600 mm, and an average annual temperature of 13.7 °C. Winter wheat is the predominant crop in the area, which is planted in late October and harvested by early June. The geographical location of the study area and the field sampling points are shown in Figure 1.

2.2. Experimental Design

Field experiments were conducted during the winter wheat heading stage in two consecutive years (2024 and 2025) at the same well-facilitated farmland demonstration area in Baisi Town. To ensure the consistency and comparability of multi-source remote sensing data, synchronized UAV flights were performed under clear-sky conditions.

On 29 April 2024 and 27 April 2025, at 10:00 a.m. local time (Beijing Time), two UAV platforms were simultaneously deployed in the study area. A fixed-wing UAV (Beijing, China) equipped with a MiniSAR system (Beijing, China) was used to acquire SAR data, while a DJI Matrice 350 RTK multi-rotor UAV (Shenzhen, China) carrying a multi-spectral camera (MicaSense, Seattle, WA, USA) was used to collect multi-spectral imagery.

Immediately after the UAV flights, field sampling campaigns were conducted to obtain in situ measurements of soil water content, vegetation water content, and other relevant biophysical parameters. The geographic coordinates of each sampling site were recorded using a real-time kinematic (RTK) GPS system (Shanghai, China) with centimeter-level accuracy. After data collection, UAV sensor data were preprocessed, and ground samples were processed through drying, weighing, and statistical analysis to obtain accurate measurements. The processed datasets were subsequently divided into training and validation sets, where the training dataset was used for modeling and the validation dataset was used for model evaluation. The RTK coordinates were then used to co-register the field measurements with the geometrically corrected UAV imagery (both multi-spectral and SAR), ensuring that each sampling point was precisely matched to the corresponding image pixels. The overall spatial registration error was evaluated to be within 0.1 m.

This study integrated UAV multi-spectral data and MiniSAR remote sensing data, combining an improved water cloud model (IWCM) with machine learning algorithms to develop a soil water content estimation framework for winter wheat fields that couples physically based constraints with data-driven modeling. The overall technical workflow is illustrated in Figure 2.

(1)

Data acquisition: A UAV equipped with a multi-spectral sensor captured spectral data in five bands: blue, green, red, near-infrared, and red edge. At the same time, the MiniSAR system collected microwave backscatter data under four polarization modes: VV, VH, HV, and HH. Field measurements of soil water content and vegetation water content were conducted simultaneously.

(2)

Data preprocessing: Multi-spectral imagery underwent mosaicking, radiometric calibration, and geometric correction. MiniSAR data were processed through geometric correction and resampling. Based on the multi-spectral imagery, six vegetation indices (e.g., NDVI and EVI) were calculated, and vegetation cover was derived.

(3)

IWCM construction: Vegetation cover was incorporated into the classical water cloud model (WCM) to construct the improved water cloud model (IWCM). This model separates the scattering contributions of vegetation and soil to the microwave signal, enabling the retrieval of polarized soil backscattering coefficients with clear physical interpretation.

(4)

Model development: Four feature input schemes were designed, including spectral band reflectance, vegetation indices, MiniSAR polarimetric parameters, and multi-source integrated variables. Two machine learning algorithms, back propagation neural network (BPNN) and random forest (RF), were employed to construct soil water content inversion models.

(5)

Model evaluation: Using field-measured soil water content data as a reference, the inversion accuracy of each model was compared and evaluated to identify the optimal model configuration.

2.3. Data Acquisition and Preprocessing

2.3.1. UAV Multi-Spectral Data

Remote sensing data were collected using a DJI Matrice 350 RTK UAV equipped with a Rededge-MX multi-spectral camera (as shown in Figure 3), which captures five spectral bands: blue (475 nm), green (560 nm), red (668 nm), near-infrared (NIR, 842 nm), and red edge (717 nm). Data were acquired from local time 11:00 to 14:00 on cloud-free days, when the air temperature remained steady and neither strong wind nor rainfall occurred. After radiometric calibration and geometric correction using Pix4D mapper 4.5.6 software, the final spatial resolution of the imagery reached 0.1 m.

2.3.2. UAV MiniSAR Data

The MiniSAR data were acquired using a fixed-wing UAV equipped with domestically developed L-band miniature synthetic aperture radar as shown in Figure 4 (MiniSAR; Model D3031, Sandia National Laboratories, USA). The system operates at a center frequency of 1.2 GHz, with a sampling rate of 600 MHz and a pulse repetition frequency (PRF) of 700 Hz. Fully polarimetric measurements (VV, VH, HV, and HH) were collected to characterize surface and vegetation scattering properties.

The MiniSAR data processing workflow consisted of preprocessing, calibration, noise suppression, and geometric correction. Initial preprocessing included image mosaicking and system calibration. Polarimetric calibration was performed using a point target-based approach with trihedral corner reflectors and active radar calibrators deployed within the study area. Calibration parameters and the system noise matrix were estimated from the measured point target responses and applied to correct the scattering matrix.

Speckle noise was reduced through multi-look processing based on sub-aperture incoherent averaging, while out-of-band noise in the range direction was removed using a finite impulse response (FIR) filter. Radiometric calibration was conducted using corner reflectors of known radar cross section (RCS). Theoretical RCS values were calculated, and a pixel integration method was applied to derive the radiometric correction coefficient for backscatter normalization.

Given the relatively large incidence angle (78°), cosine squared normalization was applied to minimize angular effects on backscatter intensity. Geometric correction was subsequently performed through co-registration with a reference optical image using homologous ground control points in Global Mapper, resulting in georeferenced SAR imagery suitable for subsequent analysis.

Based on system bandwidth and sampling parameters, the nominal range resolution was 0.25 m. After 8:1 azimuth decimation, the azimuth sampling interval was adjusted to match the range resolution. The effective spatial resolution, evaluated using the 3 dB bandwidth method, was approximately 0.6 m. The final mosaicked MiniSAR imagery achieved a spatial resolution of 0.6 m and was resampled as required for subsequent applications.

2.3.3. Field Measurement Data

Field sampling was conducted promptly after the UAV flight on the same day. A total of 55 and 90 sampling sites were established during the two campaigns, respectively. No rainfall or irrigation events occurred immediately before or after the sampling dates, ensuring stable surface water content conditions during data acquisition.

Soil water content samples were collected from the 0–10 cm layer using a soil auger in combination with the cutting ring method. Gravimetric soil water content was determined in the laboratory using the standard oven drying method at 105 °C until a constant weight was achieved. Volumetric soil water content was subsequently calculated based on measured bulk density. Vegetation parameters included plant height and vegetation water content (VWC). Plant height was measured in situ at each sampling location. For VWC determination, above-ground biomass samples were harvested and weighed immediately to obtain fresh weight (F₁). The samples were then oven-dried to a constant weight to determine dry biomass (F₂). Vegetation water content was calculated as (F₁ − F₂)/F₁ × 100%.

All field measurements were conducted synchronously with UAV overpasses to ensure temporal consistency between ground observations and remote sensing data.

Figure 5 and

Table 1. Statistics of measured SWC at sampling points.

Year	Sample Size	Min	Max	Mean	SD	95% CI Lower	95% CI Upper
2024	55	0.1744	0.3720	0.2645	0.0493	0.2512	0.2778
2025	90	0.1537	0.3583	0.2293	0.0377	0.2214	0.2372

show the descriptive statistical results of the measured data of soil water content over two years. It can be seen that the soil water content in 2024 is higher than that in 2025 as a whole.

2.4. WCM and Improved WCM

The WCM proposed by Ulaby et al. based on radiative transfer theory [32] is expressed as:

$${\sigma }_{\mathrm{c}}^0(\theta )={\sigma }_v^0(\theta )+{\gamma }^2(\theta ){\sigma }_{\mathrm{s}}^0(\theta )$$

(1)

Among these,

$${\sigma }_{\nu }^0\left(\theta \right)=A\times m_{\nu }\mathrm{c}\mathrm{o}\mathrm{s}\left(\theta \right)(1-{\gamma }^2(\theta \left)\right)$$

(2)

$${\gamma }^2\left(\theta \right)=\mathrm{e}\mathrm{x}\mathrm{p}(-2\tau \times \mathrm{s}\mathrm{e}\mathrm{c}(\theta \left)\right)$$

(3)

where ${\sigma }_{\mathrm{c}}^0(\theta )$ is the total backscattering coefficient from the vegetation-covered surface; ${\sigma }_{\nu }^0(\theta )$ is the backscattering coefficient of the vegetation layer; ${\sigma }_{\mathrm{s}}^0(\theta )$ is the backscattering coefficient of the underlying soil surface; ${\gamma }^2(\theta )$ is the two-way attenuation factor through the vegetation layer; $A$ is an empirical coefficient dependent on crop type and radar frequency, with values referenced from

Table 2. Empirical coefficients A and B for the WCM.

Parameter	All Land Uses	Rangeland	Winter Wheat	Pasture
A	0.0012	0.0009	0.0018	0.0014
B	0.0910	0.0320	0.1380	0.0840

; $\theta$ is the radar incidence angle, taken as 78° in this study; and $\tau$ is the vegetation optical thickness, expressed in terms of vegetation water content ($VWC$) as $m_{\nu }$:

$$\mathrm{\tau }=b\times m_v$$

(4)

where the value of $b$ depends on the crop type, with specific values provided in Table 2.

Based on extensive experiments, Bindlish et al. [36] summarized the empirical coefficients $A$ and $B$ for the WCM under different vegetation cover types, as shown in Table 2.

The $m_{\nu }$ was estimated by fitting the normalized difference vegetation index (NDVI). The formulas are calculated as follows:

$$m_v=0.7439\times NDVI+0.1122$$

(5)

Under actual natural conditions, the growth of winter wheat is unevenly distributed. Therefore, it is necessary to introduce a vegetation cover parameter to improve the traditional WCM. The improved WCM is expressed as follows:

$${\sigma }_{\mathrm{c}}^0(\theta )=f_v({\sigma }_v^0(\theta )+{\gamma }^2(\theta ){\sigma }_{\mathrm{s}}^0(\theta )+(1-f_v){\sigma }_{\mathrm{s}}^0(\theta ))$$

(6)

In this study, the dimidiate pixel model was employed to calculate the vegetation cover of winter wheat in the study area. The formula is expressed as follows:

$$f_v=(NDVI-NDV{I}_{\min })/(NDV{I}_{\max }-NDV{I}_{\min })$$

(7)

In the formula, $NDVI$ represents the normalized difference vegetation index and $NDV{I}_{\max }$ and $NDV{I}_{\min }$ denote the NDVI values corresponding to complete vegetation cover and exposed bare soil, respectively.

2.5. Machine Learning Model

The BPNN has been extensively employed in soil water content retrieval via remote sensing owing to its distinct advantages [21]. This approach demonstrates powerful nonlinear mapping capabilities, enabling automatic feature learning and extraction without relying on explicit mathematical formulations. It further exhibits satisfactory generalization performance and robustness against noise.

This work adopted a three-layer neural network structure comprising an input layer, a hidden layer, and an output layer. The number of neurons in the input layer was set according to the input feature scheme, while the size of the hidden layer was optimized using empirical Equation (8). The output layer contained a single neuron, corresponding to the estimated soil water content value. Key hyperparameters included the use of a logistic sigmoid (logsig) activation function, a maximum of 1000 training epochs, a learning rate of 0.01, and a training error goal of 1 × 10⁻⁶. These configurations collectively established a BPNN-based inversion model for soil water content estimation.

To enhance model reliability, all input variables were normalized prior to training. Following standard machine learning conventions, 70% of the samples were allocated for training and the remaining 30% for testing.

$$h=\sqrt{m+n}+a$$

(8)

Among these, $h$ represents the number of hidden layer nodes, $m$ and $n$ denote the number of input-layer and output-layer nodes, respectively, and $a$ is an adjustment constant between 1 and 10.

RF, first introduced by Breiman et al. [37], is a statistical ensemble learning algorithm. During training, it grows a collection of decision trees, each fitted on a bootstrap-resampled subset of the original data. At every node, only a random portion of the available features is evaluated for splitting. This two-fold source of stochasticity promotes diversity among individual trees and strengthens the overall predictive stability of the model.

As a hybrid approach combining decision trees with the bagging (bootstrap aggregating) framework, RF operates by having each tree generate predictions based on its own decision rules. The final output is obtained by aggregating the predictions of all individual trees, typically through averaging in regression tasks, thus achieving the overall regression objective of the RF algorithm.

The RF regression model demonstrates strong adaptability to complex datasets and is capable of effectively and accurately analyzing nonlinear, multivariate, and interactive effects within the data. It requires no prior assumptions regarding the relationship between independent and dependent variables or the mathematical form of the model, contributing to its broad applicability. The parameter setting of random forest in this study is shown in

Table 3. Parameter setting of random forest.

Parameter	Value Setting
Number of decision trees	100
Minimum number of leaves	3
Maximum depth per tree	70
Minimum number of observations per leaf	3
Method	Regression

.

2.6. Parameter Selection and Model Evaluation

2.6.1. Optimal Selection of Vegetation Indices

The spectral vegetation indices derived from crop canopies are rich in plant water information and are widely utilized for monitoring soil water content. Accordingly, this study selected six widely used vegetation indices—NDVI, EVI, RVI, DVI, SAVI, and TVI (as listed in

Table 4. Values of vegetation parameters used in the WCM.

Vegetation Indices	Formula
Normalized difference vegetation index (NDVI)	${\displaystyle \frac{{\rho }_{NIR}-{\rho }_R}{{\rho }_{NIR}+{\rho }_R}}$ [38]
Enhanced vegetation index (EVI)	${\displaystyle \frac{2.5\times ({\rho }_{NIR}-{\rho }_R)}{{\rho }_{NIR}+6\times {\rho }_R-7.5\times {\rho }_B+1}}$ [39]
Ratio vegetation index (RVI)	${\rho }_{NIR}/{\rho }_R$ [40]
Difference vegetation index (DVI)	${\rho }_{NIR}-{\rho }_R$ [41]
Soil adjusted vegetation index (SAVI)	${\displaystyle \frac{1.5\times ({\rho }_{NIR}-{\rho }_R)}{{\rho }_{NIR}+{\rho }_R+0.5}}$ [42]
Triangular vegetation index (TVI)	$60\times ({\rho }_{NIR}-{\rho }_G)-100\times ({\rho }_R-{\rho }_G)$ [43]

All vegetation index formulas are cited from their original sources as indicated.

)—to construct a machine learning model for soil water content inversion.

2.6.2. Feature Variable Input Settings

To compare the predictive performance of features derived from different remote sensing sources, we configured four input schemes for the machine learning models: (1) spectral reflectance bands, (2) vegetation indices, (3) MiniSAR polarimetric variables, and (4) an integrated set of all features (

Table 5. Comparison of feature input schemes.

Input Schemes	Input Variable	Output Parameters
Band spectral reflectance	Blue, Green, Red, NIR, Red Edge	SWC
Vegetation indices	NDVI [38], EVI [39], RVI [40], DVI [41], SAVI [42], TVI [43]	SWC
MiniSAR polarimetric variable	${\sigma }_s^{VV}$, ${\sigma }_s^{VH}$, ${\sigma }_s^{HV}$, ${\sigma }_s^{HH}$ [44]	SWC
All feature variables	All the variables mentioned above	SWC

). In this study, R², RMSE, MAE and 95% CI were used to evaluate the prediction effect of machine learning algorithms.

To systematically evaluate model performance, this study employed the coefficient of determination (R²), mean absolute error (MAE), root mean square error (RMSE), and the 95% confidence interval (CI) as evaluation metrics to assess model accuracy, prediction bias, and reliability.

3. Results

3.1. Soil Water Content Inversion Using BPNN with Multi-Feature Variable Input Schemes

Measured soil water content data and corresponding remote sensing features including spectral band reflectance, vegetation indices, and MiniSAR-derived polarimetric backscattering coefficients from all sampling points in the study area were randomly divided into training and testing sets at a ratio of 7:3. Using the four feature sets outlined in Table 5 as inputs, BPNN-based inversion models were trained and subsequently applied to the testing set to estimate soil water content. The agreement between predicted and observed soil water content values under different feature combinations for the two growing seasons is illustrated in Figure 6 and Figure 7.

As shown in Figure 6, the model incorporating all feature variables demonstrated good predictive performance for winter wheat during the heading stage in 2024, with an R² of 0.682 and MAE and RMSE values of 0.0298 cm³/cm³ and 0.0394. Models using only vegetation indices or MiniSAR polarimetric variables also performed satisfactorily, with R² values exceeding 0.500. In contrast, the model based solely on band reflectance variables exhibited limited performance, with an R² of 0.250, an MAE of 0.0354 cm³/cm³, and an RMSE of 0.447.

Under consistent phenological stage and inversion methodology, results from the 2025 heading season (Figure 7) revealed that the model using only band reflectance variables achieved high accuracy, with an R² of 0.679, an MAE of 0.0169 cm³/cm³, and an RMSE of 0.0217, markedly outperforming its 2024 counterpart. Models incorporating vegetation indices, MiniSAR polarimetric variables, and all features also performed satisfactorily in 2025, with R² values all above 0.500.

Overall, the BPNN algorithm combined with multiple feature variables demonstrates strong predictive capability for soil water content inversion under most conditions. However, the substantial discrepancy in performance of the band-reflectance-only model between years, such as its poor accuracy in 2024 versus its high accuracy in 2025, suggests that the robustness of BPNN may be influenced by interannual variability in feature characteristics or environmental conditions.

3.2. Soil Water Content Inversion Using RF with Multi-Feature Variable Input Schemes

Similar to the procedure used in BPNN algorithms, the four feature variables were incorporated as separate input layers. Training samples were used to fit the RF regression model, enabling the construction of a soil water content inversion model. This model was subsequently applied to the test samples to obtain predicted soil water content values. In order to more succinctly show the importance of each feature variable in random forest regression and prediction, 15 feature variables were numbered in this paper, as listed in

Table 6. Numbered name of the feature variable.

Numbered Name	Feature Variable
x1	Red
x2	Green
x3	Blue
x4	NIR
x5	Red Edge
x6	NDVI
x7	EVI
x8	RVI
x9	DVI
x10	SAVI
x11	TVI
x12	${\sigma }_s^{VV}$
x13	${\sigma }_s^{VH}$
x14	${\sigma }_s^{HV}$
x15	${\sigma }_s^{HH}$

. Figure 8 and Figure 9, respectively, show the error curve and feature importance diagram for two years when using the random forest algorithm and inputting four feature variable schemes. It can be seen that the importance of different input variables on soil moisture prediction is quite different, which provides a basic support for the use of the random forest algorithm.

Figure 10 and Figure 11 show the consistency between the predicted and measured soil water content values under different feature variable input schemes over two years using the RF method. As shown in Figure 10, for winter wheat at the heading stage in 2024, the inversion model utilizing band reflectance variables achieved the highest predictive accuracy, with an R² of 0.868, an MAE of 0.0178, and an RMSE of 0.0210. The models incorporating vegetation indices, MiniSAR polarimetric parameters, and the full feature set also demonstrated strong performance, with R² values exceeding 0.750, and both MAE and RMSE remaining below 0.0210 and 0.0260, respectively. Consistent with these trends, the 2025 results for the same phenological stage (Figure 11) showed that the model based on band reflectance variables again exhibited high accuracy, with an R² of 0.865, an MAE of 0.0152, and an RMSE of 0.0197. The models using other feature combinations, vegetation indices, polarimetric variables, and all features also performed reliably, with R² values above 0.550.

3.3. Comprehensive Analysis of Soil Water Content Retrieval

Table 7. Evaluation of soil water content retrieval accuracy using machine learning with multi-feature input schemes.

Input Schemes	Year	R2	MAE/ (cm3/cm3)	RMSE/ (cm3/cm3)	95% CI Lower	95% CI Upper
Band spectral reflectance	2024	BP *: 0.250	BP: 0.0354	BP: 0.0447	0.2408	0.2748
		RF: 0.868	RF: 0.0178	RF: 0.0210	0.2330	0.2724
	2025	BP: 0.679	BP: 0.0169	BP: 0.0217	0.2212	0.2475
		RF: 0.865	RF: 0.0152	RF: 0.0197	0.2198	0.2359
Vegetation indices	2024	BP: 0.677	BP: 0.0301	BP: 0.0375	0.2468	0.2652
		RF: 0.772	RF: 0.0208	RF: 0.0252	0.2297	0.2724
	2025	BP: 0.631	BP: 0.0193	BP: 0.0233	0.2177	0.2371
		RF: 0.678	RF: 0.0197	RF: 0.0261	0.2271	0.2397
MiniSAR polarization variable	2024	BP: 0.506	BP: 0.0341	BP: 0.0406	0.2339	0.2901
		RF: 0.780	RF: 0.0207	RF: 0.0251	0.2272	0.2718
	2025	BP: 0.508	BP: 0.0201	BP: 0.0268	0.2208	0.2373
		RF: 0.570	RF: 0.0183	RF: 0.0259	0.2266	0.2427
All characteristic variables	2024	BP: 0.682	BP: 0.0298	BP: 0.0394	0.2058	0.2577
		RF: 0.857	RF: 0.0175	RF: 0.0208	0.2270	0.2723
	2025	BP: 0.620	BP: 0.0192	BP: 0.0243	0.2250	0.2423
		RF: 0.796	RF: 0.0165	RF: 0.0207	0.2207	0.2373

* BP is BPNN.

presents the accuracy evaluation of soil water content retrieval using two machine learning algorithms with multiple feature variable input strategies. Among the 16 combinations of algorithms and feature sets evaluated, the RF model coupled with band reflectance variables yielded the best performance in 2025, achieving an R² of 0.865, MAE of 0.0152 cm³/cm³, and RMSE of 0.0197. In contrast, the poorest result was observed with the BPNN using band reflectance variables in 2024, with an R² of 0.250, MAE of 0.0354 cm³/cm³, and RMSE of 0.0447.

Comparative evaluation between the two algorithms showed that RF consistently surpassed BPNN under every feature input scheme, demonstrating its superior suitability for high-accuracy soil water content prediction in winter wheat during the heading stage. Overall, the RF algorithm consistently delivered robust inversion performance across different feature inputs and years, outperforming the BPNN in both stability and predictive capability for soil water content estimation in winter wheat.

Furthermore, among the four feature input schemes, the BPNN achieved relatively better retrieval results when using vegetation indices or all features, whereas models incorporating MiniSAR polarimetric variables showed moderate performance. Band reflectance inputs exhibited considerable instability with this algorithm. Under the RF framework, band reflectance variables produced the highest accuracy, followed by the use of all features.

4. Discussion

4.1. Discussion of Retrieval Results

Numerous studies have demonstrated that there is no consistent or explicit linear or other functional relationship among spectral reflectance, vegetation indices, SAR microwave parameters, and surface soil water content in crop-covered regions. This highlights the need for data-driven approaches to construct robust inversion models for estimating soil water content. A complex and interactive relationship exists between soil water content and crop water status. Canopy leaf water content significantly influences spectral reflectance, thereby affecting vegetation indices derived from multi-spectral data. Consequently, vegetation indices calculated from multi-spectral band reflectance have become key indicators and widely adopted methods for estimating soil water content under vegetation cover [18,20]. However, research remains limited on establishing direct correlations between soil water content and either raw band reflectance or SAR backscatter coefficients at multiple polarizations, particularly with airborne MiniSAR data. The multi-polarization capabilities of airborne MiniSAR systems offer rich and diverse information about land surface features. Therefore, integrating soil backscattering coefficients derived from vegetation scattering models into machine learning-based retrieval frameworks offers considerable promise for improving soil water content estimation accuracy.

Based on multi-spectral and MiniSAR remote sensing data acquired by UAV combined with an IWCM, four feature sets for soil water content prediction were constructed: spectral band reflectance, vegetation indices, MiniSAR polarimetric soil backscatter coefficients, and a combination of all variables. Four input schemes incorporating different combinations of these variables were designed. Soil water content inversion models were subsequently constructed using both BPNN and RF algorithms. Comparative analysis revealed that most machine learning algorithms and variable input schemes achieved satisfactory inversion performance. When employing the RF algorithm, each input scheme outperformed the corresponding BPNN model; RF models are generally more robust for small-sized to medium-sized datasets and are less sensitive to parameter initialization compared with neural networks, and thus their soil water content inversion accuracy is enhanced. Specifically, for models using MiniSAR polarimetric variables, both machine learning algorithms achieved R² values above 0.500.

The same model parameters were used for both years. When vegetation indices, a MiniSAR polarization variable, and all characteristic variables were used as input variables, the retrieval effect in 2024 was better than that in 2025 under the two algorithms (BPNN and RF). When band spectral reflectance was used as the input variable, the inversion effect in 2025 was better than that in 2024. This result may also be related to the distribution of measured data over two years and the division of the training set and the validation set. These differences may also be associated with variations in vegetation growth conditions, canopy water content, soil surface roughness, and environmental observation conditions between the two years. Although the number of samples differs between years, the model performance does not show a consistent dependence on sample size, suggesting that other factors such as vegetation water content and observation conditions play a larger role.

When comparing the results with other studies, the R², MAE, and RMSE metrics were significantly improved compared to univariate linear regression models developed in similar research contexts and crop growth stages [8,9]. These results demonstrate the feasibility and effectiveness of integrating multi-feature remote sensing inputs with machine learning algorithms for estimating surface soil water content. In addition, compared with similar studies conducted in the same region (Hebi City) [45,46], the results obtained in this study show improved performance. Some of the retrieval results based on UAV remote sensing data combined with BPNN and RF algorithms show noticeably better performance than those based on satellite remote sensing data and the support vector machine regression algorithm. At the same time, compared with the simple use of UAV multi-spectral data to construct a vegetation index for random forest algorithm inversion, this study still has certain advantages, showing the potential of multi-source remote sensing data to characterize soil moisture.

4.2. Application Prospects in Precision Agriculture

In practical agricultural applications, flight planning for UAV-based remote sensing of soil water content should be dynamically adjusted according to the phenological stages of winter wheat and real-time meteorological conditions. Based on operational monitoring requirements in the study area, the jointing and heading stages identified as critical water-sensitive periods are recommended as priority windows for UAV monitoring. Depending on soil water content dynamics, one or more flights may be conducted during each key growth period, whereas flight frequency can be reduced during other stages to enhance monitoring efficiency and cost-effectiveness. In this study, a single UAV mission covered approximately 20,000 mu of farmland, with preprocessing and generation of soil water content thematic products completed within two days, demonstrating strong timeliness and potential for operational deployment.

Furthermore, based on actual operational expenditures from the Henan experimental area, this study evaluates the economic feasibility of the UAV soil water content monitoring system. The total cost comprises hardware costs, including depreciation of UAV platforms and sensor payloads, along with consumables such as batteries and service costs, covering flight operation, data processing, and analysis. Long-term routine monitoring can distribute fixed costs across multiple missions, supporting the scalable and regular operational use of UAV remote sensing for soil water content monitoring.

In terms of system integration and practical application, the spatial distribution maps of soil water content retrieved in this study can be integrated with real-time meteorological data and crop water requirements to generate variable-rate irrigation prescription maps. These maps can be directly linked to intelligent field irrigation systems for precise implementation. This process establishes a closed-loop framework of “UAV remote sensing monitoring–soil water content assessment–irrigation decision-making field execution” offering a reliable technical pathway for precise irrigation management, water conservation, and yield enhancement in winter wheat cultivation.

Beyond agricultural applications, the proposed method may also have broader potential for environmental monitoring. For example, high-resolution soil water content information derived from UAV-based multi-source remote sensing could support ecological restoration assessment, evaluation of soil water conservation capacity, and studies related to climate change.

5. Conclusions

This study utilized UAV-based MiniSAR and multi-spectral remote sensing data. These datasets were combined with ground-measured soil moisture samples and the improved water cloud model (IWCM) to construct multiple feature input schemes for soil water content inversion. The performance of two machine learning algorithms, BPNN and RF, was compared and evaluated for inverting soil water content during the heading stage of winter wheat. The main conclusions are as follows:

(1)

Based on airborne MiniSAR data, the IWCM considering vegetation cover was employed to estimate soil water content beneath the winter wheat canopy during the same growth period over two consecutive years, deriving soil backscattering coefficients under four polarizations (VV, VH, HV, and HH). Simultaneously, six vegetation indices were extracted from multi-spectral data to characterize crop canopy conditions. Based on these variables, four feature input schemes were designed: spectral band reflectance, vegetation indices, MiniSAR polarimetric variables, and a combination of all variables. Soil water content inversion models were then constructed using BPNN and RF algorithms, respectively.

(2)

The results indicate that both machine learning algorithms achieved satisfactory prediction performance under most input schemes, validating the effectiveness of the proposed parameter configurations and feature input strategies. Among all models, the random forest model using spectral band reflectance variables performed best, with an R² of 0.865, MAE of 0.0152 cm³/cm³, and RMSE of 0.0197.

(3)

For each feature input scheme, the RF algorithm consistently outperformed the BPNN algorithm, demonstrating stronger robustness and generalization capability. This finding highlights the suitability of the RF algorithm for efficient soil water content estimation in winter wheat fields. Furthermore, the input scheme combining all feature variables achieved the best inversion performance, indicating that fusing multi-source features can better capture the variability of soil water content beneath the vegetation canopy.

Overall, this study demonstrates the feasibility of integrating multi-source remote sensing features with machine learning algorithms for soil moisture inversion. However, certain limitations remain. Besides crop spectral and SAR backscattering characteristics, other factors that may influence soil water content variation, such as canopy temperature and plant height, have not been fully considered. Future research should incorporate additional multi-source data and extend the inversion analysis to multiple soil depths and different crop growth stages to better evaluate the generalization capability of machine learning-based inversion models. Moreover, the current approach primarily relies on machine learning to fit statistical relationships between features and soil water content. Future work will consider introducing a physics-informed neural network (PINN) model, embedding physical constraints such as vegetation scattering models and soil water balance equations into the loss function of the neural network. This ensures that the model’s output not only fits the data but also adheres to physical laws, thereby improving the accuracy of soil moisture inversion. At the same time, the application of this method in yield estimation, forestry monitoring, and other fields should be further expanded to adapt to more application scenarios.