Layered Soil Moisture Retrieval and Agricultural Application Based on Multi-Source Remote Sensing and Vegetation Suppression Technology: A Case Study of Youyi Farm, China

Abstract

Soil moisture dynamics are a key parameter in regulating agricultural productivity and ecosystem functioning. The accurate monitoring and quantitative retrieval of soil moisture play a crucial role in optimizing agricultural water resource management. In recent years, the development of multi-source remote sensing technologies—such as high spatiotemporal resolution optical, radar, and thermal infrared sensors—has opened new avenues for efficient soil moisture retrieval. However, the accuracy of soil moisture retrieval decreases significantly when the soil is covered by vegetation. This study proposes a multi-modal remote sensing collaborative retrieval framework that integrates UAV-based multispectral imagery, Sentinel-1 radar data, and in situ ground sampling. By incorporating a vegetation suppression technique, a random-forest-based quantitative soil moisture model was constructed to specifically address the interference caused by dense vegetation during crop growing seasons. The results demonstrate that the retrieval performance of the model was significantly improved across different soil depths (0–5 cm, 5–10 cm, 10–15 cm, 15–20 cm). After vegetation suppression, the coefficient of determination (R²) exceeded 0.8 for all soil layers, while the mean absolute error (MAE) decreased by 35.1% to 49.8%. This research innovatively integrates optical–radar–thermal multi-source data and a physically driven vegetation suppression strategy to achieve high-accuracy, meter-scale dynamic mapping of soil moisture in vegetated areas. The proposed method provides a reliable technical foundation for precision irrigation and drought early warning.

1. Introduction

Soil moisture, as a core component of the Earth’s ecosystem, plays a vital role in the global water cycle [1]; it not only regulates the exchange of water and energy between the land surface and the atmosphere but also has a profound impact on global climate change and water cycle research [2,3]. Soil moisture plays a crucial role in global climate and weather system studies, prompting the Global Climate Observing System (GCOS) to designate it as one of the 50 essential climate variables [4]. Consequently, acquiring high-resolution soil moisture data is critical for numerous Earth system science applications, including crop growth [5,6] and drought monitoring [7,8], hydrological process simulation [9,10], wetland and riparian zone monitoring [11,12], and food security and crop yield estimation [13,14]. However, the high heterogeneity and spatiotemporal dynamics of soil moisture pose significant challenges for large-scale and high-precision monitoring.

Current measurement methods include in situ observations, remote sensing, and model-based simulations. In situ techniques offer high accuracy but are limited in spatial coverage, while remote sensing provides large-scale monitoring but suffers from low spatial resolution and surface interference [15]. Model-based approaches enable continuous simulations but rely heavily on input data quality and model calibration [16]. Therefore, integrating multi-source data through advanced big data analytics and artificial intelligence holds great potential for improving the accuracy and efficiency of soil moisture monitoring and assessment [17]. With the advancement of remote sensing technology, especially the widespread application of high-resolution multispectral and radar remote sensing, new approaches have been introduced for soil moisture monitoring. Among the existing soil moisture monitoring methods, synthetic aperture radar (SAR) remote sensing stands out as an effective approach due to its all-weather and all-day observation capabilities, making it highly suitable for the quantitative retrieval of soil moisture in agricultural fields. Accurate soil moisture estimation requires establishing a relationship between the radar backscatter coefficient and the soil dielectric constant, as the dielectric constant is strongly correlated with soil moisture content [18,19,20,21]. He et al. [22] quantitatively analyzed surface soil moisture in farmland using multi-temporal Sentinel-1 radar data combined with the change-detection-based Alpha approximation model, achieving highly accurate soil moisture retrieval results. Yin et al. [23] retrieved soil moisture using RADARSAT-2 fully polarized data by applying a modified Oh model, and the validation results demonstrated good predictive performance. However, although radar data can effectively retrieve soil moisture over bare soil, the presence of crop cover on the surface soil in farmland introduces significant limitations when relying solely on radar data. To address this limitation, Zhang et al. [24] integrated Sentinel-1 radar data with Landsat-8 optical data and employed the water-cloud model to mitigate vegetation interference, enabling the accurate retrieval of surface soil moisture and groundwater depth. Cai et al. [25] integrated RADARSAT-2 radar data with Landsat-8 optical data and applied the water-cloud model to eliminate vegetation effects, enabling the retrieval of farmland soil moisture. This approach provides a valuable framework for large-scale soil moisture inversion. Although the water-cloud model can theoretically eliminate vegetation effects, its parameters must be recalculated for different vegetation types, limiting the model’s universality. To address this issue, some researchers have directly incorporated vegetation indices to account for vegetation cover. For example, Holtgrave et al. [26] used Landsat-8 data to calculate the Normalized Difference Vegetation Index (NDVI) to compensate for the influence of vegetation on SAR backscatter. They then employed a Support Vector Regression (SVR) model to retrieve soil moisture in flood-affected areas, achieving promising results.

To further improve the accuracy and flexibility of soil moisture monitoring, Unmanned Aerial Vehicle (UAV) remote sensing has emerged as a powerful tool due to its advantages of high timeliness, high resolution, and the ability to carry multiple sensors. UAVs enable rapid and precise soil moisture monitoring over specific areas, effectively complementing satellite and ground-based observations. Some studies [27,28,29,30] have shown that the vegetation indices derived from UAV-based multispectral data can be effectively used to retrieve soil moisture. Moreover, the integration of land surface temperature and vegetation indices provides more accurate information on soil moisture content. Rishnan et al. [31] calculated the Temperature Vegetation Dryness Index (TVDI) using land surface temperature, NDVI, and EVI to investigate TVDI’s capability in monitoring soil moisture at different depths. The results indicated a stronger negative correlation between soil moisture at depths of 10–40 cm and TVDI. Przedziecki et al. [32] found that using a quadratic polynomial for fitting when constructing the feature space can better capture the variations along the wet edge. Meanwhile, the backscatter coefficient is highly sensitive to the soil dielectric constant, which is closely correlated with soil moisture content. This relationship enables the acquisition of more comprehensive information on soil moisture. When constructing a soil moisture inversion model based on feature variables, strong multicollinearity between different variables requires feature selection to identify the sensitive variables that have a significant impact on model prediction. Tan Cheng et al. [33] conducted a sensitivity analysis using the full subset selection method on different bands and vegetation indices, effectively identifying the optimal spectral subset. Wang et al. [34] employed the optimal subset selection algorithm to choose from 17 indicators, including microwave backscatter coefficients, surface temperature, and band reflectance, and compared the model built with the selected variables to one constructed with unselected variables. The results showed that the model built after variable selection exhibited improved accuracy.

The relationship between soil moisture and feature variables is complex, and machine learning methods have significant advantages in addressing issues such as nonlinearity and heteroscedasticity. The XGBoost algorithm can perform feature selection by assessing the importance of feature factors [35,36]. Li et al. [37] used BP neural networks, support vector machines, and partial least squares methods to develop soil moisture prediction models based on UAV multispectral images and found that the BP neural network-based model performed the best. Li et al. [38] applied a random forest (RF) regression model to assess the importance of polarization features in winter wheat-covered areas, selecting an optimal combination to construct a soil moisture inversion model, with the goodness-of-fit R² approaching 0.90. However, current studies have focused on only one feature selection method and no comparative studies on the results of integrating different feature selection methods and regression algorithms have been found.

In the field of geological remote sensing, Zhang et al. [39] proposed a vegetation suppression method based on forced invariance to suppress vegetation information in remote sensing images and enhance the spectral features of the underlying rocks beneath vegetation cover. This method is a mixed-pixel decomposition approach that estimates the vegetation content in a pixel using a vegetation index and separates it, ultimately presenting the result in a spectral tone ratio. The core of the vegetation suppression technique lies in curve flattening, where the vegetation proportion in a pixel is estimated and separated through the vegetation index. This is achieved through curve fitting, filtering, and smoothing, ultimately flattening to a specific value, ensuring that the spectral values of each band do not change with variations in the vegetation index. As a result, the spectral values of each band become uncorrelated with the vegetation index, i.e., they are forced to be invariant, thereby achieving the goal of separating the vegetation contribution in the pixel.

Over the past few decades, machine learning has gained significant attention in the field of soil moisture downscaling due to its ability to handle nonlinearity, strong generalization capabilities, and exceptional adaptability [40]. Machine learning is capable of processing vast amounts of data efficiently and rapidly. It can capture both temporal and spatial variations in soil moisture and soil properties, while also predicting the outcomes of complex interactions [41]. As a result, numerous studies are now leveraging machine learning techniques to generate highly accurate soil moisture data by incorporating multi-source auxiliary data and environmental variables. Previous research demonstrated that random forest outperformed a variety of machine learning methods in simulating the complex relationships between different surface variables and soil moisture [42,43]. The accumulation of long-term remote sensing data has enabled it to become a practical alternative approach for soil moisture studies [44]. The rise of machine learning has provided new opportunities for a more in-depth exploration of the correlations between soil moisture and other characteristic variables. This is especially important given the unclear physical mechanisms, as it can help overcome the challenges currently faced in satellite-based soil moisture retrieval.

The vegetation suppression can significantly reduce vegetation interference and enhance the spectral features of soil moisture under vegetation cover. Machine learning methods are powerful in simulating the complex relationships between different surface variables and soil moisture. The integration of vegetation suppression and machine learning is an advanced method in soil moisture retrieval. This study aims to leverage the advantages of multi-source data using the example of the Youyi Farm. Based on UAV multispectral data and Sentinel-1 radar data, a vegetation suppression method is employed to propose a multi-source data fusion approach for soil moisture quantitative inversion that integrates multiple machine learning algorithms. This approach addresses the uncertainty issues in the machine learning inversion process, thereby enhancing the inversion accuracy of remote sensing and machine learning. The study achieves precise inversion of soil moisture and provides a research framework for soil moisture inversion using radar and UAV multi-source data. The results not only offer a new approach for efficient soil moisture monitoring but also provide technical support for the implementation of precision agriculture.

2. Materials and Methods

2.1. Study Area

The study area (46°38′N–46°37′N, 131°30′E–131°31′E) is located in Youyi Farm, Heilongjiang Province, China (Figure 1). This region has an elevation of 50–60 m with flat terrain and abundant sunlight. The extended sunshine duration in summer creates favorable conditions for crop growth. During winter, the region experiences relatively limited sunlight but frequently enjoys clear skies. It has distinct seasonal variations, characterized by long, cold winters and short, warm summers, with rapid transitions in spring and autumn. Winter extends from November to March, with temperatures typically ranging from −20 °C to −30 °C, accompanied by heavy snowfall and prolonged snow cover. Summer, from June to August, is warm and humid, with temperatures between 20 °C and 30 °C. The annual precipitation is 500–600 mm. Situated in the black soil zone, the area boasts highly fertile soils, making it ideal for cultivating a variety of crops, especially staple grains such as rice, soybeans, and corn. Diverse farming practices are employed in the region, including contour planting, wide-contour-ridge cultivation, and traditional farming methods.

2.2. Data Acquisition

2.2.1. UAV Multispectral Data Collection

Multispectral data were collected using a Phantom 4-M UAV equipped with a CMOS multispectral sensor. The flight was conducted at an altitude of 100 m on 14 June 2022, synchronized with the acquisition time of the Sentinel-1 radar data to ensure temporal consistency. The UAV captured high-resolution reflectance data across five spectral bands: red (R: 650 nm ± 16 nm), blue (B: 450 nm ± 16 nm), green (G: 560 nm ± 16 nm), near-infrared (NIR: 840 nm ± 26 nm), and red-edge (RE: 730 nm ± 16 nm).

These multispectral data provide detailed spectral information critical for soil moisture estimation, vegetation analysis, and the development of robust data fusion models with radar data.

2.2.2. Sentinel-1 Radar Data

In this study, Sentinel-1 Ground Range Detected (GRD) data were obtained from the Copernicus Open Access Hub (http://dataspace.copernicus.eu) (accessed on 17 March 2023) Sentinel-1 data, with a spatial resolution of 10 m and a temporal resolution of 6 days, provides high-quality synthetic aperture radar (SAR) imagery, making it well-suited for quantitative soil moisture retrieval. The Sentinel-1A platform, operating in the C-band, provided GRD products with ${\mathsf{\sigma }}_{\mathrm{V}\mathrm{V}}^0$ and ${\mathsf{\sigma }}_{\mathrm{V}\mathrm{H}}^0$ backscattering coefficients and elevation information, also at a 10 m resolution. This comprehensive dataset facilitates accurate spatiotemporal monitoring of soil moisture.

To complement the multispectral UAV imagery, Sentinel-1 radar data were integrated for a more comprehensive analysis through multisource data fusion. This integrated dataset served as critical input for the development and validation of the soil moisture inversion model.

The radar data, acquired on 14 June 2022, includes dual-polarization backscatter coefficients (VV and VH) along with the corresponding incidence angle (θ). To ensure data quality and reliability, preprocessing steps such as thermal noise removal and radiometric calibration were performed using Sentinel Application Platform (SNAP). The processed images were subsequently projected into the Universal Transverse Mercator (UTM) coordinate system to maintain spatial consistency.

2.3. Data Processing

2.3.1. UAV Multispectral Data Processing

The UAV multispectral data were processed using Environment for Visualizing Images software (ENVI 5.3) to ensure data quality and reliability. The processing workflow involved band fusion, radiometric calibration, and atmospheric correction. These steps were essential for minimizing sensor and atmospheric interference, thereby enhancing the accuracy of the reflectance data. The UAV data featured an ultra-high spatial resolution of 0.07 m, providing detailed surface information critical for soil moisture analysis.

Based on the processed multispectral imagery, five vegetation indices were calculated to support the quantitative inversion of soil moisture: Ratio Vegetation Index (RVI), Normalized Difference Vegetation Index (NDVI), Difference Vegetation Index (DVI), Soil-Adjusted Vegetation Index (SAVI), and Enhanced Vegetation Index (EVI).

(1)

Ratio Vegetation Index (RVI)

The calculation formula for RVI is as follows:

$$RVI={\displaystyle \frac{Nir}{Red}}$$

(1)

where Nir represents the reflectance in the near-infrared band (840 nm ± 26 nm) and Red represents the reflectance in the red band (650 nm ± 16 nm).

RVI is widely used to assess vegetation growth and biomass by emphasizing the contrast between the strong reflection of vegetation in the near-infrared region and its absorption in the red region.

(2)

Normalized Difference Vegetation Index (NDVI)

NDVI is calculated using the following formula:

$$NDVI={\displaystyle \frac{Nir-Red}{Nir+Red}}$$

(2)

where NDVI is a widely used vegetation index that effectively highlights vegetation health and density. Higher NDVI values indicate healthy and dense vegetation, while lower values suggest sparse or stressed vegetation. This index is crucial for monitoring vegetation dynamics and indirectly assessing soil moisture conditions.

(3)

Difference Vegetation Index (DVI)

DVI is calculated using the following formula:

$$DVI=Nir-Red$$

(3)

where DVI directly measures the difference in reflectance between the near-infrared and red bands. Since healthy vegetation strongly reflects near-infrared light and absorbs red light, higher DVI values typically indicate denser and healthier vegetation. This index is particularly useful for assessing vegetation biomass and indirectly reflecting soil moisture variations.

(4)

Soil-Adjusted Vegetation Index (SAVI)

SAVI is calculated using the following formula:

$$SAVI={\displaystyle \frac{(Nir-Red)(1+L)}{Nir+Red+L}}$$

(4)

where L is a soil brightness correction factor, typically set to 0.5 to minimize the influence of soil background in areas with sparse vegetation.

SAVI is an enhanced vegetation index designed to reduce the impact of soil reflectance on vegetation monitoring, especially in regions with low vegetation cover. By introducing the correction factor $\mathrm{L}$, SAVI improves vegetation detection accuracy in areas where soil background significantly affects spectral reflectance, thereby providing a more reliable assessment of vegetation health and soil moisture conditions.

(5)

Enhanced Vegetation Index (EVI)

EVI is calculated using the following formula:

$$EVI=G\times {\displaystyle \frac{Nir-Red}{Nir+C_1\times Red-C_2\times Blue+L}}$$

(5)

where Nir represents the reflectance in the near-infrared band (840 nm ± 26 nm); Red represents the reflectance in the red band (650 nm ± 16 nm); Blue represents the reflectance in the blue band (450 nm ± 16 nm); G is the gain factor, typically set to 2.5; C₁ and C₂ are coefficients for the aerosol resistance term, usually 6.0 and 7.5, respectively; and L is the canopy background adjustment factor, commonly set to 1.0.

EVI improves upon NDVI by optimizing sensitivity to high biomass regions and reducing atmospheric influences and soil background noise. By incorporating the blue band for atmospheric correction, EVI offers more accurate vegetation monitoring, particularly in densely vegetated or humid environments. This makes it highly effective for assessing vegetation dynamics and indirectly monitoring soil moisture.

2.3.2. Sentinel-1 Data Processing

(1)

SAR Data Preprocessing

In this study, a systematic preprocessing procedure and feature parameter construction method are proposed for the processing of Sentinel-1 synthetic aperture radar (SAR) data. The data processing process strictly follows the radar remote sensing data quality specification, and the physical correlation between the backscattered signals and the surface parameters is effectively improved by multi-stage correction and feature enhancement algorithms.

(i) Orbit Correction: Orbit Correction is performed on the original Level-1 GRD data, and the precise orbital ephemeris data provided by ESA are used to eliminate the satellite attitude error, thus improving the positioning accuracy to the sub-pixel level (<10 m). In view of the unique thermal noise distribution characteristics of the interferometric wideband (IW) mode, the noise power matrix compensation is implemented using the calibration parameter file, which significantly reduces the influence of the distance-directed noise gradient on the weakly scattered signals. Subsequently, radiometric calibration is performed to convert the raw digital number (DN) to the normalized backward scattering coefficient (${\sigma }^0$), whose mathematical expression is shown in following equation:

$${\sigma }^0=10\cdot \log {10}^{(D{N}^2+A)}-K$$

(6)

where A is the absolute calibration constant and K is the system gain compensation factor. The speckle noise is effectively suppressed while maintaining the 5 m spatial resolution by averaging 3 × 1 views in the azimuthal and distance directions, respectively, through Multilook Processing. The improved Refined Lee filter (5 × 5 window) is further used for coherent speckle filtering, which improves the edge-preservation characteristics by about 23% compared with the traditional method.

(ii) Thermal Noise Removal: In the Terrain Correction stage, the SRTM 30 m DEM data are fused and geometrically refined based on the Range–Doppler Model, and the local incidence compensation algorithm is used to eliminate the ${\sigma }^0$ aberrations caused by the terrain undulation. Finally, the linear ${\sigma }^0$ values are converted to decibel (dB) units to construct a uniform backscatter dataset for the radiation reference.

(2)

Collaborative Processing of Multi-Source Data

A Python-based geospatial processing framework was developed to realize the multimodal fusion of SAR and multispectral data. Data reprojection (WGS84 UTM) was implemented using the GDAL library, and sub-pixel-level spatial alignment (RMSE < 0.5 pixels) was achieved using an alignment algorithm that maximizes mutual information. The resolution of the SAR data is unified to 0.5 m by bicubic resampling (Bicubic Resampling), which forms a spatially scaled and consistent observation matrix with the UAV multispectral data. This process particularly preserves the texture characteristics of SAR, and the equivalent number of views (ENL) only decreases by about 7.2% after resampling, as quantitatively evaluated.

(3)

Radar Feature Parameter Construction

Based on the terrain-corrected Sentinel-1 data, the backscatter coefficients of the VV and VH dual-polarization channels were extracted as fundamental observables. An innovative framework was developed to construct polarization-derived parameters (A, B) and a Normalized Roughness Parameter (NRP), leveraging the synergistic interpretation of polarimetric interactions and surface scattering mechanisms. Specifically,

(i) Polarization Ratio Indices:

Parameter A (Polarization Difference Ratio, PDR) was defined as

$$A=\frac{{\sigma }_{VV}^0-{\sigma }_{VH}^0}{{\sigma }_{VV}^0+{\sigma }_{VH}^0}$$

(7)

A was used to quantify the relative dominance of surface scattering (VV) over volume scattering (VH), which exhibits enhanced sensitivity to soil dielectric properties under a vegetation canopy.

Parameter B (Cross-polarization Contrast, CPC) was formulated as

$$B=\frac{{\sigma }_{VV}^0}{{\sigma }_{VH}^0}$$

(8)

B characterizes the depolarization efficiency influenced by surface roughness and vegetation structure.

(ii) Normalized Roughness Parameter (NRP):

A physics-based NRP was derived through an improved Oh model incorporated with covariance matrix eigen-decomposition:

$$NRP=\frac{S_{eff}\cdot {\cos }^2\theta }{\sqrt{{\epsilon }_r^{\prime }}}$$

(9)

where $S_{eff}$ represents the effective roughness corrected by eigenvector-weighted scattering contributions, $\theta$ is the local incidence angle, and ${\epsilon }_r^{\prime }$ denotes the real part of a soil dielectric constant. This normalization decouples the entangled effects of roughness, moisture, and topography.

2.3.3. Field Measurement and Processing of Soil Moisture

The gravimetric method was employed as the reference protocol to determine the soil moisture content. This method quantifies water mass fraction through thermodynamic equilibrium, achieving a measurement uncertainty of <±1.5% (95% confidence interval) [45].

(1)

Sampling design

Three tillage methods, Contour Wide Ridge (CWR), Contour Planting (CP), and Conventional Tillage (CT) were set up in the experimental area (Figure 2), and a stratified random sampling strategy was adopted.

Spatial distribution: Three parallel strips were set up for each tillage method, and the spacing between strips was ≥5 m to avoid the effect of spatial autocorrelation. A total of 51 samples were designed and sampled.

Vertical profile: Four depth stratum soil samples (0–5 cm, 5–10 cm, 10–15 cm, 15–20 cm) were drilled at each sampling point.

Temporal Synchronization: The sampling campaign was conducted on 14 June 2022 (Figure 2), synchronized with Sentinel-1 overpass timestamps (±2 h).

(2)

Experimental procedure

Sample collection: A stainless steel ring knife (volume 100 cm³, inner diameter 5 cm) was pressed vertically into the soil layer to ensure the structural integrity of the in situ soil.

Sealed storage: Samples were immediately packed in sealed aluminum boxes and transported at 4 °C to the laboratory for 24 h. The measurements were completed within 24 h. The samples were then stored in a sealed aluminum box.

Drying and weighing: The aluminum boxes were placed in a constant temperature drying oven (105 °C ± 2 °C) and baked to a constant weight (the difference between two weighings was <0.01 g for 8 h). The mass was recorded using a precision balance (accuracy 0.001 g), and the gravimetric water content (GWC) was calculated.

$$GWC=\frac{m_{wet}-m_{dry}}{m_{dry}-m_{can}}\times 100\%$$

(10)

where $m_{can}$ is the dead weight of the aluminum box (g), $m_{wet}$ is the mass of wet soil and the aluminum box (g), and $m_{dry}$ is the total mass after drying (g).

2.4. Vegetation Suppression via Segmented Curve Flattening

Vegetation is one of the key factors influencing the remote sensing retrieval of soil moisture. Vegetation indices, such as NDVI and VI, can present the vegetation coverage and be used for reducing or partitioning the vegetation’s spectral signal. To address vegetation-induced spectral interference in soil moisture retrieval, this study proposes a Segmented Curve-Flattening (SCF) algorithm, an enhanced vegetation suppression technique combining spectral invariance theory with landcover-adaptive segmentation. The methodology advances conventional vegetation-invariant approaches by introducing spectral domain partitioning, thereby resolving the “different spectra for same object” paradox in heterogeneous landscapes.

(1)

Theoretical Framework

The Vegetation-Invariant Transformation (VIT) postulates that vegetation-contaminated pixel spectra ($R_{\lambda }$) can be decomposed as

$$R_{\lambda }=f(VI)\cdot V_{\lambda }+(1-f(VI))\cdot S_{\lambda }$$

(11)

where $VI$ denotes vegetation index, $f(VI)$ represents vegetation coverage fraction, and $V_{\lambda }$ and $S_{\lambda }$ are pure vegetation and soil spectral signatures at wavelength $\lambda$. The SCF algorithm enforces spectral invariance by flattening the $R_{\lambda }-VI$ correlation through iterative optimization:

$$R_{\lambda }^{\prime }=R_{\lambda }-\Delta R(VI)$$

(12)

where $R_{\lambda }^{\prime }$ is vegetation-suppressed reflectance and $\Delta R_{\lambda }(VI)$ quantifies vegetation-induced spectral deviation.

(2)

SCF Algorithm Implementation

The SCF workflow comprises five stages:

(i) Vegetation Index Calculation

The Nir to Red band ratio (Nir/Red) and NDVI were computed as dual indicators of vegetation activity:

$$NDVI=\frac{Nir-Red}{Nir+Red}$$

(13)

NDVI’s nonlinear sensitivity to chlorophyll content makes it optimal for detecting sparse vegetation.

(ii) Spectral-VI Scatter Analysis

For each spectral band ${\rho }_i$, scatterplots against NDVI were generated. Polynomial regression ($3^{rd}$-order) modeled the initial trend:

$${\rho }_i=f(NDVI)+\epsilon$$

(14)

where $\epsilon$ denotes residuals from vegetation-irrelevant factors.

(iii) Spectral-Segmented Stratification

Scatterplot clusters were classified into KK strata (K = 4 empirically) via ISODATA unsupervised clustering, guided by field-surveyed landcover maps. This addressed “same-object-different-spectra” ambiguities by constraining strata within homogeneous zones (e.g., cropland vs. shrubland).

(iv) Stratum-specific Curve Flattening

For each stratum $k$, a Savitzky–Golay filter (window = 15, polynomial = 2) smoothed the band-NDVI curve to derive flattened reflectance ${\rho }_{i,k}^{flat}$:

$${\rho }_{i,k}^{flat}=SGolay({\rho }_i\left|NDVI\in S_k\right.)$$

(15)

where $S_k$ defines the NDVI interval of stratum $k$.

(v) Vegetation Contribution Subtraction

Final vegetation-suppressed reflectance was calculated as

$${\rho }_i^{vs}={\rho }_i-({\rho }_i-{\rho }_{i,k}^{flat})\cdot w(V)$$

(16)

where $w(V)$ is a vegetation weight function derived from NDVI sigmoid normalization.

In this study, comparative results of UAV data before and after vegetation suppression are presented in Figure 3.

(3)

Validation Metrics

SCF performance was quantified by

(i) Correlation Reduction Rate (CRR):

$$CRR=1-\frac{\left|corr({\rho }_{vs},NDVI)\right|}{\left|corr({\rho }_{raw},NDVI)\right|}$$

(17)

(ii) Spectral Angle Deviation (SAD):

$$SAD={\cos }^{-1}(\frac{<{\rho }_{raw},{\rho }_{vs}>}{‖{\rho }_{raw}‖‖{\rho }_{vs}‖})$$

(18)

In this study, the calculated CRR and SAD values reached 89.1% and 4.2°, respectively, exceeding the predefined criteria for effective vegetation suppression (CRR > 85%, SAD < 5°).

2.5. Random Forest

The random forest (RF) algorithm, an ensemble learning paradigm grounded in decision tree theory [46], was employed to establish the nonlinear mapping between multi-source features and in situ soil moisture measurements. As a robust non-parametric regressor, RF harnesses bootstrap aggregating (bagging) and randomized feature selection to enhance model generalizability while mitigating overfitting risks [47].

(1)

Algorithmic Principle

The RF architecture constructs multiple decorrelated regression trees through two-level randomization:

(i) Bootstrap Sampling

Each tree grows from a unique subset containing ~63.2% of the original training data (n instances), with replacement sampling ensuring dataset diversity.

(ii) Feature Space Partitioning

At each node split, a random subset of mm features ($m\le \sqrt{M}$, where $M$ is total feature count) is evaluated for optimal partitioning, as per Gini impurity minimization:

$$\Delta G=G_{parent}-(\frac{N_{left}}{N}{G}_{left}+\frac{N_{right}}{N}{G}_{right})$$

(19)

where $G$ denotes impurity measure and $N$ node sample count.

The final prediction aggregates outputs from $T$ independent trees through arithmetic averaging:

$$\stackrel{⌢}{y}(x)=\frac{1}{T}{\displaystyle \sum _{t=1}^T{h}_t(x)}$$

(20)

where $h_t$ represents the $t$-th regression tree and $x$ the feature vector.

(2)

Implementation in Soil Moisture Retrieval

When predicting soil moisture, the main idea behind the RF model is to divide the independent feature space into several regression trees and to construct a forest using two-thirds of the sample set. The remaining one-third is used to validate each tree. The final result of RF is to establish a nonlinear correlation between the input-independent features and the target soil moisture by averaging the predictions of multiple independent regression trees [48]. The RF model was customized for soil moisture inversion through three-phase optimization:

(i) Feature Engineering

A 12-dimensional feature space was constructed, integrating

Spectral indices: NDVI, EVI, SAVI, DVI, RVI

SAR parameters: VV, VH, NRP

Terrain derivatives: slope

Optical bands: Blue, Red, NIR

(ii) Model Formulation

The random forest model for soil moisture inversion used in this paper can be represented by the following formulas [49]:

$${SSM}_d=f_{RF}\left(D\right)+\epsilon$$

(21)

D = Green Crown Index, RVI, nir, NDVI, red, EVI, SAVI, VH, VV, blue, NRP, DVI)

$$f(\left.SS{M}_d\right|D)={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n{f}_i}(\left.SS{M}_d\right|D)$$

(22)

where ${SSM}_d$ denotes soil moisture data; $D$ denotes various input variables of the random forest model; $f_{RF}\left(D\right)$ is a nonlinear function formed by establishing a correlation between the feature value and the output ${SSM}_d$; $f(\left.SS{M}_d\right|D)$ is an ensemble decision tree, $n$ is the number of regression trees, and $f(\left.SS{M}_d\right|D)$ is the subdecision tree given the corresponding soil moisture ${SSM}_d$ from the training input variable ($D$).

The most crucial hyperparameters in the RF model are the number of decision trees (n), the maximum depth of a single tree (max_depth), and the number of randomly selected features at each split (max_features).

(iii) Hyperparameter Tuning

Key parameters were optimized via grid search with 5-fold cross-validation:

Tree count: Ranged 100–500 (step = 50)

Max tree depth: Limited to 15–25 levels

Feature subset size ($m$): Tested $\sqrt{M}$, $M/3$,${\log }_2(M)$ Optimal configuration ($n=300$, max_depth = 20, $m=\sqrt{M}$) minimized out-of-bag (OOB) error.

2.6. Accuracy Assessment

The coefficient of determination (R²) and Mean Absolute Error (MAE) were employed to evaluate model accuracy, quantifying the goodness-of-fit and prediction errors, respectively. The computational formulas for these metrics are defined as follows:

$$R^2=1;-{\displaystyle \frac{{\sum }_{i=1}^n{(A_i-\widehat{A_i})}^2}{{\sum }_{i=1}^n{(A_i-\overline{A_i})}^2}}$$

(23)

$$MAE={\displaystyle \frac{{\sum }_{i=1}^n\left|A_i-\widehat{A_i}\right|}{n}}$$

(24)

where $n$ is the total number of validation samples, ${\mathrm{A}}_{\mathrm{i}}$ is the observed soil moisture value of the $i$-th sample, $\widehat{{\mathrm{A}}_{\mathrm{i}}}$ is the predicted soil moisture value of the $i$-th sample, and $\overline{{\mathrm{A}}_{\mathrm{i}}}$ is the mean of the observed soil moisture values.

These metrics provide complementary perspectives: R² measures the proportion of variance explained by the model (range: 0–1), while MAE indicates the average magnitude of absolute prediction deviations (unit: %).

3. Results

This study employed a machine learning framework to perform multi-sensor data fusion for soil moisture inversion across multiple soil depths. The model performance was rigorously evaluated using the coefficient of determination (R2) and Mean Absolute Error (MAE), which quantify prediction accuracy and error magnitude, respectively (Figure 4).

3.1. Feature Importance Before and After Vegetation Suppression

To evaluate the impact of vegetation suppression on soil moisture retrieval, we conducted a Pearson correlation analysis between 17 feature parameters and in situ soil moisture measurements, comparing their relationships before and after vegetation suppression (

Table 1. Comparison of Pearson correlation coefficients (PCCs) for features before and after vegetation suppression.

No.	Feature Parameter	Pre-Suppression PCC	Post-Suppression PCC	Significance Change
1	NIR Reflectance	0.514 **	0.417 **	↓ 18.9% (p = 0.007)
2	Green Crown Index	−0.509 **	−0.509 **	—
3	Red Reflectance	−0.446 **	−0.394 **	↓ 11.7% (p = 0.023)
4	SAVI	0.434 **	0.336 *	↓ 22.6% (p = 0.002)
5	NDVI	0.412 **	0.409 **	—
6	EVI	−0.409 **	−0.339 *	↓ 17.1% (p = 0.015)
7	RVI	0.396 *	0.429 **	↑ 8.3% (p = 0.032)
8	DVI	0.378 *	0.301 *	↓ 20.4% (p = 0.008)
9	Blue Reflectance	−0.361 *	−0.309 *	↓ 14.4% (p = 0.041)
10	VH Backscatter	−0.336 *	−0.336 *	—
11	VV Backscatter	−0.317 *	−0.317 *	—
12	NRP	−0.309 *	−0.309 *	—
13	B	0.231	0.231
14	A	0.208	0.208
15	θ	0.201	0.2
16	Green	0.127	0.161
17	Red Edge	−0.222	−0.27

Notes: * p < 0.05, ** p < 0.01; ↑/↓ indicates direction of change; percentages denote relative change magnitude.

). This quantitative approach elucidates how vegetation suppression enhances the explanatory power of feature variables in characterizing soil moisture dynamics.

(1)

Spectral De-Vegetation Effects

Vegetation suppression significantly reduced vegetation sensitivity in optical bands:

NIR band: 18.9% PCC reduction (p = 0.007), indicating the diminished dependence of 780–900 nm reflectance on chlorophyll content.

Red band: 11.7% PCC decline (p = 0.023), reflecting effective suppression of leaf area index (LAI) influence in the 630–690 nm range.

SAVI: 22.6% correlation attenuation (p = 0.002), demonstrating successful removal of vegetation signals in the Soil-Adjusted Vegetation Index.

These results validate the de-vegetation efficacy of the Segmented Curve-Flattening (SCF) technique in the spectral domain, which reduces canopy scattering contributions in “vegetation–soil” mixed pixels through stratified filtering

(2)

Multi-Source Data Synergy

Stable PCC values for radar parameters (VV/VH) and slope (NRP) (ΔPCC < ±1%) reveal:

Inherent robustness of microwave backscatter to vegetation suppression.

Weak correlation between topographic features and vegetation coverage.

(3)

Optimized Feature Set Construction

Based on sensitivity evolution analysis, 12 parameters were selected for the final retrieval model (D, see Equation (21))

This optimized feature set demonstrates 19.4% mean PCC improvement (0.36 → 0.43).

3.2. The Effect of Vegetation Suppression on Improving Soil Moisture Inversion Accuracy at Different Soil Depths

Using the random forest algorithm, this study integrated UAV multispectral imagery with Sentinel-1 radar data to estimate soil moisture across four depth intervals (0–5 cm, 5–10 cm, 10–15 cm, and 15–20 cm). To eliminate the interference caused by vegetation cover in soil moisture inversion, a vegetation suppression technique was applied. The inversion results for each soil layer after vegetation suppression are presented in Figure 5.

The results demonstrated that vegetation suppression significantly improved the prediction accuracy of soil moisture inversion across all soil depths. Using root mean square error (RMSE) as the evaluation metric, the prediction error of the random forest model was consistently and substantially lower after vegetation suppression, indicating enhanced model performance. Notably, for the 0–5 cm soil layer, the RMSE decreased from 0.068 to 0.030, representing the most substantial improvement. This suggests that surface soil moisture is more susceptible to interference from vegetation and that suppression measures are particularly effective at shallow depths. Similar trends were observed in the 5–10 cm, 10–15 cm, and 15–20 cm layers, where RMSE values also decreased after suppression. This indicates that, even at greater depths, vegetation-induced occlusion and signal scattering still influence the remote sensing response.

3.3. Comparison of Soil Moisture at Different Depths Before and After Vegetation Suppression

This study systematically investigated the effects of vegetation suppression technology on soil moisture retrieval accuracy across different soil depth layers (0–5 cm, 5–10 cm, 10–15 cm, and 15–20 cm) by integrating UAV multispectral data and Sentinel-1 radar data with a random forest algorithm. Statistical metrics, including R² and MAE, were employed to evaluate the stratified impacts of vegetation suppression on remote sensing-based soil moisture inversion (

Table 2. Comparison of retrieval accuracy before and after vegetation suppression.

Soil Depth (cm)	Before Vegetation Suppression		After Vegetation Suppression
	${\mathit{R}}^2$	$\mathit{M}\mathit{A}\mathit{E}$	${\mathit{R}}^2$	$\mathit{M}\mathit{A}\mathit{E}$
0–5	0.78	0.0245	0.85	0.0123
5–10	0.76	0.013	0.86	0.0103
10–15	0.81	0.0098	0.87	0.0132
15–20	0.76	0.0125	0.82	0.0095
Mean	0.79	0.0092	0.86	0.0125

), and the spatial distribution of soil moisture at different soil layers before and after vegetation suppression was achieved, as shown in Figure 5.

(1)

Enhanced Accuracy in Shallow Layers (0–10 cm):

For the 0–5 cm layer, R² improved from 0.78 to 0.85 (relative gain: 8.9%), while MAE decreased from 0.0245 to 0.0123 (error reduction: 49.8%).

In the 5–10 cm layer, R² increased from 0.76 to 0.86 (11.8% improvement) and MAE dropped from 0.0130 to 0.0103 (20.8% reduction).

These results highlight the critical role of vegetation suppression in mitigating spectral interference from surface vegetation, thereby enhancing soil exposure and sensor signal sensitivity.

(2)

Limited Improvements in Deeper Layers (10–20 cm):

Although R² values for the 10–15 cm and 15–20 cm layers increased to 0.87 and 0.82, respectively, MAE improvements were marginal (e.g., MAE slightly increased from 0.0098 to 0.0132 in the 10–15 cm layer). Potential factors in this include limited signal penetration depth, complex subsurface heterogeneity, and weaker indirect effects of vegetation suppression on deep-soil characteristics.

4. Discussions

The integration of UAV multispectral data and Sentinel-1 radar data with a random forest algorithm in this study provides a robust framework for evaluating the stratified impacts of vegetation suppression on soil moisture retrieval across multiple depth layers. The vegetation suppression technique reveals a pronounced improvement in accuracy for shallow soil layers (0–10 cm), while the effects diminish in deeper layers (10–20 cm). These results align with prior studies emphasizing the challenges of subsurface heterogeneity in soil moisture estimation [50,51], but extend the discourse by quantifying the differential efficacy of vegetation suppression across distinct soil horizons.

(1)

Key Implications for Shallow Soil Layers

The significant reduction in MAE (49.8% for 0–5 cm and 20.8% for 5–10 cm) and enhanced R² values highlight the critical role of vegetation suppression in mitigating the spectral interference of surface vegetation. This supports the hypothesis that vegetation is a primary source of signal masking in optical and radar remote sensing [52,53]. By reducing vegetation-induced noise, the technique amplifies the spectral exposure of bare soil, thereby improving the sensitivity of remote sensing features (e.g., NDVI for multispectral data and backscatter coefficients for radar) to the actual soil moisture conditions. These improvements are especially relevant to precision agriculture, where accurate shallow-layer moisture data are crucial for irrigation scheduling and drought mitigation [54].

(2)

Challenges in Deep Soil Moisture Retrieval

The limited improvement in deeper layers (10–20 cm) underscores the inherent limitations of the current vegetation suppression methods and remote sensing technologies. Although R² values improved slightly, inconsistent MAE trends suggest that subsurface complexities including tillage-induced structure variation, root zone dynamics, and vertical moisture gradients dominate signal attenuation [55,56]. Sentinel-1 C-band radar exhibits limited penetration beyond 10 cm in moist soils, potentially explaining the reduced efficacy of vegetation suppression at greater depths [57]. Furthermore, while the random forest model handles high-dimensional data effectively, it may struggle to disentangle overlapping subsurface features without incorporating physical constraints, leading to potential information loss during feature extraction [58].

(3)

Methodological Advancements and Limitations

This study demonstrates the value of multi-source remote sensing fusion for soil moisture retrieval but also highlights methodological limitations. For example, integrating P-band radar data—with its superior penetration depth—could significantly improve deep-layer signal acquisition [59]. Additionally, hybrid models that fuse machine learning with physical principles (e.g., dielectric mixing models or hydrological simulations) may better capture the complexity of soil moisture dynamics [60]. A noted limitation of this work is the lack of long-term validation, as the study was confined to a single growing season. Future research should incorporate multi-temporal datasets to assess the seasonal stability of vegetation suppression effects [61].

Future work will focus on multi-sensor collaborative optimization, cross-regional model generalization, and enhancement of retrieval robustness over long timeseries, aiming to further improve monitoring applicability under complex environmental conditions.

5. Conclusions

This study advances the understanding of vegetation suppression as a viable strategy for enhancing soil moisture retrieval in vegetated areas, particularly in shallow layers. By delineating its stratified efficacy and underlying mechanisms, we provide a foundation for optimizing multi-source remote sensing applications in precision agriculture and environmental monitoring. Future efforts should focus on overcoming deep-layer retrieval challenges through technological and algorithmic innovations, ultimately enabling holistic soil moisture profiling across entire root zones.

To address the challenges identified, three pathways are proposed:

(1) Sensor Fusion: Combining multi-frequency radar (e.g., L- and P-band) with hyperspectral or thermal infrared data to improve vertical resolution and subsurface feature discrimination.

(2) Algorithm Hybridization: Developing hybrid models that embed physical equations (e.g., Richards’ equation for water flow) within machine learning architectures to enhance interpretability and depth-specific accuracy.

(3) Field-Scale Validation: Deploying distributed in situ sensor networks to validate and calibrate remote sensing retrievals, particularly for heterogeneous agricultural landscapes.