Influence of Soil Background Noise on Accuracy of Soil Moisture Content Inversion in Alfalfa Fields Based on UAV Multispectral Data

Abstract

Soil moisture plays a critical role in the global water cycle, the exchange of matter and energy within ecosystems, and the movement of water in plants. Accurate monitoring of soil moisture is essential for drought early warning systems, irrigation decision-making, and crop growth assessment. The use of drone-based multispectral remote sensing technology for estimating the soil moisture content offers advantages such as wide coverage, high accuracy, and efficiency. However, the soil background can often interfere with the accuracy of these estimations. In specific environments, such as areas with strong winds, removing soil background noise may not necessarily enhance the precision of estimates. This study utilizes unmanned aerial vehicle (UAV) multispectral imagery and employs a vegetation index threshold method to remove soil background noise. It systematically analyzes the response relationship between spectral reflectance, spectral indices, and the soil moisture content in the top 0–10 cm layer of alfalfa; constructs K-Nearest Neighbors (KNN), Random Forest Regression (RFR), ridge regression (RR), and XG-Boost inversion models; and comprehensively evaluates model performance. The results indicate the following: (1) The XG-Boost model validation set had the highest R² value (0.812) when spectral reflectance was used as the input variable, which was significantly better than the other models (R² = 0.465 to 0.770), and the RFR model validation set had the highest R² value when the spectral index was used as the input variable (0.632), which was significantly better than the other models (R² = 0.366 to 0.535). (2) After removing soil background noise, the accuracy of the soil moisture estimates for each model did not show significant changes; specifically, the R² value for the XG-Boost model decreased to 0.803 when using spectral reflectance as the input, and the R² value for the RFR model dropped to 0.628 when using spectral indices. (3) Before and after removing the soil background noise, the spectral reflectance can provide more accurate data support for the inversion of the soil moisture content than the spectral index, and the XG-Boost model is the most effective in the inversion of the soil moisture content when using the spectral reflectance as the input variable. The research findings provide both theoretical and technical support for the retrieval of the surface soil moisture content in alfalfa using drone-based multispectral remote sensing. Additionally, they offer evidence that validates large-scale soil moisture remote sensing monitoring.

1. Introduction

Soil moisture is a core component of the Earth’s hydrosphere, playing a vital role in key processes such as water cycling and plant moisture transport. Its spatial and temporal distributions profoundly influence the co-evolution of the biosphere, atmosphere, and lithosphere through the flow of matter and energy [1]. As a key indicator for drought monitoring, irrigation decision-making, and crop growth assessment, accurate measurement of the soil moisture content plays a crucial strategic role in enhancing the effectiveness of drought early warning systems and optimizing irrigation management [2]. Compared with the limitations of traditional soil moisture monitoring, such as limited coverage, weak sample representativeness, and time-consuming and labor-intensive acquisition, UAV remote sensing technology, with its advantages of a wide monitoring range, high spatial accuracy, low implementation costs, and high operational efficiency, has gradually become the core means of dynamically monitoring the soil water content and has provided more efficient and comprehensive technological support for grasping changes in soil moisture in real time.

Unmanned aerial vehicle (UAV) remote sensing mainly acquires soil surface spectral properties, thermal radiation, and electromagnetic wave signals using sensors such as hyperspectral, multispectral, thermal infrared, synthetic aperture radar (SAR), and electromagnetic wave sensors and realizes soil moisture content inversion by combining these data with physical models or machine learning algorithms [3,4]. Compared to other sensors, multispectral sensors are not affected by temperature, ensuring high data stability and easy data acquisition. They can directly reflect the differences in the spectral reflectance of surface features. Future research could fuse multispectral and thermal infrared remote-sensing data to construct multidimensional feature inputs, thereby substantially improving the accuracy of soil moisture estimation and the robustness of predictive models.

In recent years, numerous scholars have conducted in-depth research on the refinement of data processing technology, innovation and optimization related to inversion models, accuracy validation and uncertainty analysis, and crop type expansion. Research on improving the refinement of data processing technology mainly involves efficient rejection of soil background interference and accurate extraction of key features (spectral reflectance, texture features, temporal features, and multi-source features) [5,6]. The SHAP method utilizes cooperative game theory to quantify the contribution of each feature to the prediction outcomes, thereby enhancing the model’s interpretability. Applying the SHAP method to soil moisture prediction research enables quantitative identification of the contributions of input variables to the soil moisture content, providing a reference for variable selection when certain input factors are missing. Fu et al. [7] introduced the SHAP method to quantitatively assess the contributions of input variables in ensemble learning models. Innovation and optimization related to inversion models include the development from empirical models to the fusion of mechanistic and data-driven models, covering physical models of spectral/temperature signals and the soil water content based on radiative transfer/heat conduction mechanisms. The most commonly applied statistical methods are the traditional machine learning models (Random Forest Regression (RFR), Support Vector Machine (SVM), and K-Nearest Neighbors regression (KNN)) and the deep machine learning models (Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs)) [8,9]. Random Forest Regression (RFR), XG-Boost, K-Nearest Neighbors (KNN), and ridge regression (RR) have all demonstrated good performance in UAV remote sensing inversion [10,11]. Compared to Support Vector Machines (SVMs), Convolutional Neural Networks (CNNs), and Recurrent Neural Networks (RNNs), models like RFR, XG-Boost, KNN, and RR offer key advantages, including higher computational efficiency, stronger inherent interpretability, better generalization capabilities and robustness against overfitting for medium-sized datasets, a relatively simple hyperparameter tuning process, and generally superior modeling efficiency and competitiveness in structured data prediction tasks.

Accuracy validation and uncertainty analysis mainly systematically evaluate inversion reliability through ground-truthing data validation, multi-model comparison (cross-validation, with an independent validation set comparing the generalization ability), and uncertainty source analysis (sensor error, meteorological conditions, crop phenology, soil heterogeneity, etc.) [12]. The research covers a wide range of crop types and has gradually expanded from bulk crops such as winter wheat and maize to cash crops such as cotton, as well as saline and alkaline crops and pasture grasses in alpine regions. The above series of studies have greatly improved model universality and practical application suitability [13,14].

Soil background noise refers to additional signals received by sensors when they detect the spectral signals reflected by vegetation, which also include reflections from the soil background. Consequently, canopy reflectance encompasses the combined spectral response of both the vegetation components and the soil background. Removing the soil background is a crucial step in image processing techniques. By utilizing the spectral differences of surface features and the principles of radiative transfer, it is possible to separate the target features (such as vegetation) from the soil signals in mixed signals. This effectively reduces the effect of soil interference on the inversion of target parameters, allowing for more accurate characterization of the target features (Zhang et al. [15]). Almeida-Naunay A F [16] utilized spectral differences to calculate thresholds and implemented threshold optimization to remove soil background pixels. They found that reducing soil interference resulted in improved predictive performance. The current mainstream rejection methods include the band selection method, the vegetation index method, physical model inversion, etc. The steps include screening sensitive bands, constructing a functional relationship, and decomposing total radiance. For example, Yang et al. [17] found that using the red-to-green ratio index (RGRI) method to remove the soil background resulted in a coefficient of determination for predicting corn soil moisture that was greater than that obtained with the Otsu threshold method, while the performance without removing the soil background was the worst. Zhang et al. [18] found that by improving the vegetation index threshold method to remove the soil background of winter wheat, the effect of vegetation index inversion with removal of the soil background was always inferior to that without removal of the soil background. Da et al. [19] combined supervised classification to remove the soil background with shading and concluded that the accuracy of inversion of the maize soil water content based on the CNN, BPNN, and PLSR models was significantly improved after removal of the soil background. Wang et al. [20] used a supervised classification method to remove the soil background for corn soil moisture inversion and found that data correlation and the accuracy of the regression model were consistently lower after removal than when the soil background was not eliminated. Although existing research generally supports the idea that removing the soil background can enhance accuracy, its general applicability is still controversial. The mechanism coupling the removal operation with the input variables and the inversion model has not been fully clarified, and the studies cover relatively limited vegetation types (e.g., maize, winter wheat, cotton, etc.), so the scope of the conclusions needs to be expanded and verified. Research has found that deep learning, particularly neural networks (CNNs), can enhance agricultural functions such as soil classification, crop monitoring, and nutrient management. Additionally, Dhanya [21] employed YOLO and similar object detection algorithms to improve the agricultural image preprocessing stage by removing backgrounds and analyzing specific regions. Necati Cetin [22] utilized convolutional neural networks (CNNs) and the integrated YOLOv5 model to remove backgrounds and enhanced agricultural applications by employing precise texture classification based on texture and color features while reducing background noise interference. The deep learning-based object detection framework can dynamically predict bounding boxes and confidence scores, ensuring effective separation of soil from diversified backgrounds that include complex conditions such as shadows or debris while also facilitating data preparation and enhancing feature stability. Previous research utilizing UAV remote sensing combined with machine learning models has primarily focused on inversion of the soil moisture content for corn and winter wheat. However, studies examining the effects of alfalfa and soil background noise on the surface soil moisture content in alfalfa are relatively scarce. Therefore, this study employed UAV remote sensing technology to monitor the surface soil moisture content in alfalfa. As the ‘King of Forages’ [23], alfalfa (Medicago sativa L.) has a high nutritional value (a crude protein content of 15–25%), has strong resistance to stress (drought, cold, and salinity), and provides ecological services (rhizobacteria nitrogen fixation and fertilization, reductions in soil erosion, and improvements in the soil structure). The canopy height of alfalfa (Medicago sativa) typically ranges from 10 to 65 cm and varies significantly through different growth stages. During the peak growth period, the canopy covers a relatively large area; however, post-harvest, canopy coverage declines significantly. Additionally, factors such as intercropping combinations, seeding methods, seeding density, and environmental conditions (such as soil moisture and light intensity) influence the canopy structure of alfalfa. Gansu is the core production area of high-quality alfalfa and the focal area of the national ‘grain-to-feed’ strategy. The development of its grass industry is of strategic significance in guaranteeing the supply for regional animal husbandry and promoting the optimization of the structure of the agricultural and animal husbandry industry [24]. However, it is difficult for traditional alfalfa soil water content monitoring methods to meet the precise management needs of large-scale alfalfa planting. Based on this, this study took alfalfa as its research object; adopted UAV multispectral remote sensing technology, combined with the vegetation index threshold method, to eliminate soil background interference; systematically explored the response relationships between spectral features and the alfalfa surface soil water content; and constructed an inversion model using a machine learning algorithm. The objectives of this study were as follows: (1) We quantified the impact of the soil background on the spectrum–moisture relationship and clarified the degree of interference caused by soil noise. (2) We analyzed the effects of different combinations of variables on the enhancement of inversion accuracy by using the original spectral reflectance, the spectral index, and the culled optimized spectral features as input variables to determine the core spectral parameters affecting the moisture content. (3) based on four machine learning algorithms, namely Random Forest Regression (RFR), XG-Boost, K-Nearest Neighbors (KNN), and ridge regression (RR), we compared the inversion accuracy of the model before and after the removal of the soil background and screened the optimal model so as to provide technological support for real-time monitoring of the soil moisture content and accurate irrigation decision-making in alfalfa planting areas in Gansu.

2. Materials and Methods

2.1. An Overview of the Study Area

This experiment was conducted from April to October 2024 at the Irrigation Experiment Station of the Jingtai Chuan Electricity Lift Irrigation Water Resources Utilization Center in Gansu Province (37°12′59″ N, 104°05′10″ E, mean elevation: 1572 m, Figure 1). The soil moisture deficit in the Jingtai region significantly constrains local ecological sustainability and agricultural development, which gives this study critical scientific and practical significance. This region has a temperate continental arid climate with abundant light and sparse rainfall, and the annual average sunshine hours, frost-free period, radiation, air temperature, precipitation, and evaporation are 2652 h, 191 d, 6.18 × 10⁵ J·cm⁻², 8.6 °C, 201.6 mm, and 2761 mm, respectively (Figure 2). The soil in the test area was loamy, and its physical and chemical properties are listed in

Table 1. Physical and chemical properties of soil and meteorological data.

Index	Numeric Value	Unit
dry bulk density	1.35	g cm−3
field capacity	24.1	%
PH	8.11
organic matter	6.15	g kg−1
total nitrogen	1.58	g kg−1
total phosphorus	1.36	g kg−1
total potassium	34.16	g kg−1
fast-acting nitrogen	74.22	mg kg−1
fast-acting phosphorus	32.99	mg kg−1
fast-acting potassium	147.80	mg kg−1
precipitation	276	mm
average daily temperature	18.8	°C

.

2.2. Experimental Design

By studying the soil moisture content of alfalfa, this research can provide a scientific basis for optimizing planting management and improving water resource utilization efficiency and ecological protection. Alfalfa (Gannong 3) primarily grows from April to October. During the flowering and bud stages, canopy coverage is high and the impact of the soil background can be neglected; therefore, this study focused on the branching stage of alfalfa. The experimental field measured 27 m in length from east to west and 40 m from north to south. To enhance the universality of the inversion model in heterogeneous artificial grasslands, this study divided the area into 16 alfalfa sub-regions, with each subplot covering an area of 54 m² (6 m × 9 m). Drip irrigation was used in the study area. The drip-tape spacing was 40 cm, the emitter spacing was 30 cm, and the emitter flow rate was 2.0 L·h⁻¹. Irrigation volumes were monitored with a water meter (precision: 0.0001 m³) and regulated by ball valves. During the experiment, field management of all plots followed local conventional practices. The experimental plots were set up in a completely randomized design with four groups of irrigation gradients. (An irrigation gradient refers to the spatial variability or continuous changes in water availability or soil moisture resulting from irrigation activities. Such gradients can form naturally or can be artificially designed, with the core aspect being an uneven distribution of moisture and its resulting impacts.) Each irrigation gradient is expressed as a percentage of the soil moisture content relative to the field capacity (θf). (Field capacity refers to the maximum amount of water that soil can retain in its capillary pores after gravitational water has drained away. It is typically expressed as a percentage of the dry weight of the soil.) The field capacity levels were as follows: 45–55% θf, 55–65% θf, 65–75% θf, and 75–85% θf.

2.3. UAV Multispectral Remote Sensing Data Acquisition

2.3.1. Multispectral Remote Sensing Image Acquisition

For this experiment, remote sensing data collection was conducted on 19 April 2024, 20 April 2024, and 21 April 2024 during clear and windless weather with sufficient light. Multispectral remote sensing images were collected using a DJI Matrice 300 RTK quadcopter UAV (Shenzhen DJI Innovation Technology Co., Ltd., Shenzhen, China). (Figure 3) The UAV was equipped with an MS 600 Pro multispectral camera (Changguang Yuchen Information Technology Equipment (Qingdao) Co., Ltd., Qingdao, China) (Figure 4). The flight altitude was set to 30 m, with a forward overlap rate of 80% and a side overlap rate of 70%. The flight time was from 12:00 to 13:00. The multispectral camera lens was pointed vertically downward. Each flight followed a predetermined route, and the center wavelength, wave width, and diffuse reflector plate reflectance of each band are provided in

Table 2. Center wavelength, bandwidth, and diffuse reflector reflectance of each band.

Spectral Band	Center Wavelength (nm)	Bandwidth (nm)	Reflectance of Diffuse Reflector (%)
Blue	450	35	60
Green	555	25	60
Red	660	20	60
Red edge 1	720	10	60
Red edge 2	750	15	60
NIR	840	35	60

. The flight speed was 2.7 m/s, and images of the calibration plates were taken on the ground before and after each flight. These calibration plate images were used for reflectance calibration to compensate for variations in illumination conditions and to ensure data consistency. The ground resolution was 2.16 cm.

2.3.2. Multispectral Remote Sensing Image Processing

In this study, a total of four ground control points (GCPs) were established within the experimental area, and their coordinates were precisely determined using Real-Time Kinematic (RTK) positioning technology. Subsequently, geographic registration and radiometric correction were carried out using Pix4D Mapper version 4.8.0, and manual marking of the ground control points was employed to enhance location accuracy. The root-mean-square errors (RMSEs) of the generated digital orthophoto in the x-axis, y-axis, and z-axis directions were 0.23 m, 0.31 m, and 0.27 m, respectively, indicating that the orthophoto exhibited high spatial accuracy.

2.3.3. Removal of Soil Background Noise

During the drone data collection process, the canopy reflectance received by the sensor represented a mixed spectral response from both the vegetation and the soil background. To eliminate the interference of soil background noise, masking was performed on the multispectral images of the alfalfa and soil within each cell using a vegetation index [25] threshold in ENVI 5.3 software. Given that the NDVI values of crops are greater than 0 and the NDVI values of soil are less than 0, the NDVI method was employed for threshold segmentation, with a threshold set at 0, effectively removing soil background noise [26]. The vegetation index threshold method is based on the advantages of the vegetation index [27] in distinguishing between crop and non-crop pixels, thereby facilitating noise removal.

The vegetation index threshold method significantly outperforms supervised classification and spectral unmixing methods in terms of operational complexity and computational efficiency when it comes to removing soil background noise interference [28]. Supervised classification relies on a large number of high-quality training samples to distinguish between complex land cover categories, including various soil types. This process involves feature selection, classifier training, and optimization, resulting in high computational costs and susceptibility to sample representativeness and labeling subjectivity. In contrast, spectral unmixing requires accurate acquisition and validation of pure spectral signatures for vegetation and soil endmembers, as well as solving complex mixing models (such as linear or nonlinear models). The accuracy of this approach is highly sensitive to endmember variability and model assumptions, making the computational process relatively complex. In contrast, the vegetation index threshold method requires only an empirical or statistical determination of a single (or limited) vegetation index threshold for binary segmentation. The algorithm is straightforward, easy to implement, and highly efficient, making it particularly suitable for large areas and near-real-time processing scenarios. For areas with high vegetation cover, the threshold method can effectively distinguish between vegetation and non-vegetation. Due to its standardized process and direct results (generating binary masks), it usually provides stable and repeatable preliminary outcomes, thereby avoiding the additional errors associated with vague category definitions in supervised classification and the uncertainties of endmember extraction in spectral unmixing.

2.4. Ground Data Acquisition and Processing

While the UAV acquired the multispectral remote sensing images, alfalfa topsoil samples were manually and synchronously collected on the ground in the 16 sample areas, at a soil depth of 0–10 cm, using the five-point sampling method, and finally, the mean values of the five-point sampling were taken as the sample data. The soil samples collected at each point were put into aluminum boxes, and the soil mass water contents were measured by the drying method in an oven at a constant temperature of 105 °C for 24 h. The average values were taken as the soil water contents of the sampling points. The total number of soil samples was 48, and all samples strictly followed a normal distribution. Finally, 70% of the data were randomly selected for modeling, and 30% were selected for validation.

2.5. Spectral Index Selection and Calculation

Spectral indices can be used to extract feature information utilizing the ratios or differences between certain bands in remote sensing data. In remote sensing data processing, spectral indices are widely used in the extraction of feature information related to vegetation, soil, and water bodies. The vegetation index of the crop canopy in the experimental area was calculated by a specific algebraic combination of the spectral reflectance of each band [29]. Six relevant spectral indices were determined using the related literature, and the six spectral reflectances and six spectral indices used in this study include the red band (red, R), blue band (blue, B), green band (green, G), near-infrared band (NIR), red edge 1 band (red edge 1), red edge 2 band (red edge 2), normalized difference vegetation index (NDVI), ratio vegetation index (RVI), difference vegetation index (DVI), green index (GI), simple ratio pigment index (SRPI), and red-to-green ratio index (RGRI), which were calculated according to the equations presented in

Table 3. Spectral index formulas.

Spectral Indices	Acronyms	Formulation	Reference
Normalized difference vegetation index	NDVI	(NIR − R)/(NIR + R)	[30]
Ratio vegetation index	RVI	NIR/R	[31]
Difference vegetation index	DVI	NIR − R	[32]
Green index	GI	G/R	[33]
Simple ratio pigment index	SRPI	B/R	[34]
Red-to-green ratio index	RGRI	R/G	[35]

. Under alfalfa cover conditions, these vegetation indices are important tools for estimating the soil moisture content. In particular, the normalized difference vegetation index (NDVI) and the difference vegetation index (DVI) demonstrate significant inversion effects [8].

2.6. Soil Moisture Content Inversion Model Construction

In this study, the model input variables were divided into 2 groups: a spectral reflectance variable group and a spectral index variable group. Four machine learning algorithms, namely RFR, XG-Boost, KNN, and RR, were used to build alfalfa surface soil water content inversion models with and without the removal of soil background noise, for a total of 16 soil moisture content inversion models. Different models exhibit varying performances on the same data due to their unique algorithmic characteristics. RFR is an integrated learning algorithm that performs regression tasks by aggregating multiple decision trees, which can effectively deal with complex datasets and nonlinear relationships among features [36]. KNN is an instance-based learning algorithm for classification and regression that measures the distances between different feature values. It makes regression predictions by analyzing the nearest neighbors of the input samples and is outstanding when dealing with scenarios with complex data distributions or obscure boundaries [37]. RR, as an enhanced linear regression method, is designed to address issues of multicollinearity [38]. XG-Boost is short for Extreme Gradient Boosting. It is a machine learning algorithm implemented in the gradient boosting framework [39]. It is an efficient and flexible gradient boosting algorithm that emphasizes speed, performance, and regularization, and it is suitable for applications involving large-scale datasets and complex problems.

The four machine learning algorithms were implemented using the sklearn library in python, and the specific steps were data loading preparation and preprocessing, data partitioning, and model evaluation. Hyperparameter optimization was performed for KNN, RFR, XG-Boost, and RR using a grid search combined with 4-fold cross-validation, and the optimal parameters were selected based on the maximum mean R² value. This study systematically evaluated the predictive performance of four machine learning algorithms—KNN, RFR, XG-Boost, and RR—using 4-fold cross-validation. First, the original dataset was divided into a training set (comprising 70% of the samples) and an independent test set (comprising 30% of the samples) in a 7:3 ratio. Subsequently, 4-fold cross-validation was implemented within the training set. The training samples were evenly divided into four mutually exclusive subsets (KNN1–KNN4, RFR1–RFR4, XG-Boost1–XG-Boost4, and RR1–RR4). In each iteration, three of the folds were used as training data, while the remaining fold served as validation data. This process cycled through all four folds to obtain robust internal validation results. After completing the cross-validation, the final generalization performance of each algorithm was assessed using the test set, which had not been involved in any model tuning (KNN Test, RFR Test, XG-Boost Test, and RR Test). Predictive results were generated to ensure the objectivity and reliability of the model selection process (Figure 5).

2.7. Data Analysis

2.7.1. Pearson Correlation Analysis

Remote sensing images usually contain multiple bands (e.g., visible, near-infrared, short-wave infrared, etc.), and direct use of all bands will lead to data redundancy and an increase in computational complexity, so Pearson correlation analysis was used to screen the sensitive spectral index combinations and sensitive spectral reflectance combinations [40]. Pearson correlation analysis, as a statistical method used to measure the linear relationship between two variables, centers on the Pearson correlation coefficient, which has values ranging from −1 to 1, with 1 indicating a perfect positive correlation, −1 indicating a perfect negative correlation, and 0 indicating no correlation. The higher the absolute value of r, the stronger the linear correlation between the predictor variable and the target variable. According to the evaluation criteria for Pearson’s correlation coefficient (|r|), |r| ≥ 0.8 indicates a high correlation; 0.5 ≤ |r| < 0.8 indicates a moderate correlation; 0.2 ≤ |r| < 0.5 indicates a low correlation; and |r| < 0.2 indicates essentially no correlation [41].

2.7.2. Soil Moisture Content Inversion and Accuracy Evaluation

This study evaluated the accuracy of the soil moisture content (SMC) inversion models using three metrics: root-mean-square error (RMSE), Mean Absolute Error (MAE), and the coefficient of determination (R²). RMSE indicates the closeness of predicted values to actual values, while R² measures the goodness of fit of a model. An R² value closer to 1, an RMSE closer to 0, and a smaller MAE all indicate an inversion model with better accuracy and precision. The spectral reflectance and different spectral indices with and without soil background noise were used to invert the moisture content of alfalfa surface soil. A linear model was established with the spectral reflectance and spectral index as independent variables and the soil moisture content as the dependent variable, and the coefficients of determination (R²) and root-mean-square errors (RMSEs) of the models before and after the removal of soil background noise were compared to analyze the impact of soil background noise on the inversion of the soil moisture content.

3. Results

3.1. Analysis of Correlations Between Spectral Reflectance and Spectral Index and Soil Moisture Content

Pearson correlation coefficients were used to conduct a multifactorial correlation analysis of the independent variables (spectral reflectance and the spectral index) and the dependent variable (the SMC) (Figure 6). By selecting relevant features, the model’s generalization ability and robustness could be improved, thereby enhancing its applicability and performance across different environments, datasets, and application scenarios. Before removing the soil background noise, the correlations between the spectral reflectances and the SMC, from strongest to weakest, were red edge 750, NIR, red, blue, red edge 720, and green, among which the absolute values of the correlation coefficients of red edge 750 and NIR were greater than 0.5. After removing the soil background noise, the highest correlation was found for NIR, with an absolute correlation coefficient value of 0.7, followed by red edge 750, with an absolute correlation coefficient value of 0.69. Before and after removing the soil background noise, the correlations between red, blue, red edge 720, and green and the SMC were all lower.

From the Pearson correlations between the spectral indices and the SMC (Figure 6), it can be seen that before removing the soil background noise, the spectral indices with higher correlations with the SMC were the DVI, NDVI, and RVI, all of which had absolute correlation coefficients greater than 0.5. After removing the soil background noise, the spectral indices with higher correlations with the SMC were also the DVI, NDVI, and RVI, which all had absolute correlation coefficients exceeding 0.5. Before and after removing the soil background noise, the correlations between the SRPI and GI and the SMC were lower.

3.2. Soil Moisture Content Inversion Model Based on Spectral Reflectance

The selected spectral reflectances, red edge 750 and NIR, were used as input variables to establish soil water content inversion models using four machine learning methods, namely RFR, RR, KNN, and XG-Boost (

Table 4. The accuracy of the soil moisture content inversion models based on spectral reflectance and the spectral indices.

	Model	Soil Background	Modeling Set			Validation Set
			R2	RMSE	MAE	R2	RMSE	MAE
Spectral reflectance	KNN	unremoved	0.783	0.017	0.013	0.770	0.023	0.019
		removed	0.714	0.022	0.014	0.692	0.021	0.015
	RFR	unremoved	0.768	0.019	0.014	0.762	0.021	0.021
		removed	0.829	0.015	0.010	0.710	0.027	0.017
	RR	unremoved	0.463	0.033	0.026	0.465	0.020	0.016
		removed	0.422	0.034	0.025	0.442	0.021	0.017
	XG-Boost	unremoved	0.884	0.014	0.014	0.812	0.017	0.013
		removed	0.867	0.015	0.015	0.803	0.018	0.018
Spectral index	KNN	unremoved	0.606	0.028	0.017	0.535	0.021	0.017
		removed	0.446	0.033	0.022	0.463	0.024	0.023
	RFR	unremoved	0.772	0.020	0.017	0.632	0.020	0.018
		removed	0.737	0.020	0.018	0.628	0.026	0.021
	RR	unremoved	0.255	0.039	0.031	0.366	0.021	0.018
		removed	0.253	0.039	0.025	0.351	0.024	0.022
	XG-Boost	unremoved	0.613	0.026	0.013	0.503	0.027	0.016
		removed	0.568	0.023	0.017	0.460	0.037	0.025

). Before and after removing the soil background noise, all models except the ridge regression (RR) achieved good inversion results. K-Nearest Neighbors (KNN), Random Forest Regression (RFR), and XG-Boost demonstrated solid performance, with coefficients of determination above 0.6 and root-mean-square errors (RMSEs) below 0.1 for both the modeling and validation sets. The XG-Boost model demonstrated optimal performance. Before removing the soil background noise, the modeling set had an R² value of 0.884 and an RMSE of 0.014, while the validation set had an R² value of 0.812 and an RMSE of 0.017. After removing the soil background noise, the modeling set had an R² value of 0.867 and an RMSE of 0.015, while in the validation set the R² value reached 0.803 and the RMSE was 0.018. In summary, among the four machine learning models constructed based on spectral reflectance, XG-Boost demonstrated a strong capability to perform soil moisture content inversion. In Table 4, it can be observed that the validation results of the dataset after removing soil background noise show little variation compared to the results before the removal of soil background noise. The R² value of the validation set of the XG-Boost model decreased by 0.009, the R² value of the validation set of the RFR model decreased by 0.052, the R² value of the validation set of the KNN model decreased by 0.078, and the R² value of the validation set of the RR model decreased by 0.023.

3.3. Spectral Index-Based Soil Moisture Content Inversion Model

The selected spectral indices, the DVI, NDVI, and RVI, were used as input variables to establish RFR, RR, KNN, and XG-Boost machine learning models to predict the SMC (Table 4). Before and after removing the soil background noise, the R² values of the KNN, RFR, and XG-Boost models for both the training and validation sets were all above 0.4, while the RMSE values were all below 0.1, with the exception of the RR model. The RFR model had the best inversion effect. The R² value of the modeling set was 0.772 and the RMSE was 0.02 before removing the soil background noise, while the R² value of the validation set was 0.632 and the RMSE was 0.02. After removing the soil background noise, for the modeling set the R² value was 0.737 and the RMSE was 0.02, while for the validation set the R² value was 0.628 and the RMSE was 0.026. In summary, among the four machine learning models constructed based on the spectral indices, the RFR model demonstrates a strong ability to invert the soil moisture content.

In Table 4, the validation results for the dataset after removing soil background noise exhibit only slight differences compared to the results obtained from the dataset without noise removal. The R² value of the validation set of the XG-Boost model decreased by 0.043, the R² value of the validation set of the RFR model decreased by 0.004, the R² value of the validation set of the KNN model decreased by 0.072, and the R² value of the validation set of the RR model decreased by 0.015.

3.4. Comprehensive Evaluation of the Model

In the soil moisture content inversion model established based on spectral reflectance, the R² values of the modeling set and the validation set were mostly 0.7~0.8 before removing the soil background noise. After removing the soil background noise, the R² values of the modeling set and the validation set were also mostly 0.7~0.8 (Table 4). In the inversion model of the soil moisture content based on the spectral index, the R² values of the modeling set were mostly 0.5~0.7 and the R² values of the validation set were mostly 0.4~0.6 before removing the soil background noise. After removing the soil background noise, the R² values of the modeling set were also mostly 0.5~0.7, and the R² values of the validation set were also mostly 0.4~0.6 (Table 4). A comprehensive analysis of Table 4 revealed that the four machine learning models with spectral indices as the input variables had better soil water content inversion and better stability than those with spectral reflectances as the input variables. Furthermore, all models showed that the prediction performance after removing soil background noise had only minimal differences compared to the results obtained without noise removal. Through significance testing, all models yielded p-values lower than 0.01 (p < 0.01), indicating a high level of statistical significance for each model. This result confirms that the established models can effectively explain the variation in the dependent variable rather than attributing this variation to random error (

Table 5. Model significance testing.

	Model	Soil Background	p-Value
Spectral reflectance	KNN	unremoved	0.00002
		removed	0.0002
	RFR	unremoved	0.00003
		removed	0.00002
	RR	unremoved	0.002
		removed	0.00003
	XG-Boost	unremoved	0.000001
		removed	0.000001
Spectral index	KNN	unremoved	0.00009
		removed	0.0005
	RFR	unremoved	0.00002
		removed	0.000002
	RR	unremoved	0.00002
		removed	0.000008
	XG-Boost	unremoved	0.000003
		removed	0.000001

).

When using spectral reflectance as an input variable, the R² value of the XG-Boost model for the validation set consistently outperformed the three other inversion models both before and after the removal of soil background noise (Figure 7). In contrast, when spectral indices were used as input variables, the R² value of the Random Forest Regression (RFR) model for the validation set was greater than those of the three other models, regardless of whether soil background noise was removed (Figure 8). Therefore, in this study, the XG-Boost model was utilized with spectral reflectance as the input variable to construct an alfalfa soil moisture inversion map without removing soil background noise (Figure 9).

Residual analysis confirmed the performance of the KNN, RFR, RR, and XG-Boost models using both spectral reflectance and spectral indices as input variables. As illustrated in Figure 10 and Figure 11, the residuals exhibited a relatively random distribution across the entire range of predicted values. This indicates that the model maintains a consistent error distribution at different levels of prediction, thereby supporting the validity of the model’s assumptions. Furthermore, the residuals were randomly distributed around the zero line and did not exhibit a noticeable trend with changes in predicted values, indicating that the model does not suffer from systematic errors.

4. Discussion

4.1. The Effects of the Input Variables on the Accuracy of Soil Moisture Content Inversion

The core principle of inverting the soil moisture content based on spectral reflectance and spectral indices lies in the fact that soil moisture significantly alters the dielectric constant and surface characteristics, which in turn affect the absorption and scattering behavior of electromagnetic waves in specific wavelength bands. Notably, there are significant differences in the sensitivity of various bands to changes in the moisture content. At the same time, a spectral index constructed by multi-band synergy can further amplify the moisture signal and effectively suppress interference factors such as vegetation and soil background. Existing studies have shown that inversion models with spectral reflectance as the input variable generally have high accuracy. When Zhang et al. [42] and others used partial least squares regression (PLSR), stepwise regression, and ridge regression models to invert the surface soil moisture content, the R² values exceeded 0.7. Yang et al. [43] and others obtained similar results by PLSR and stepwise regression models. Wang et al. [44] employed a Stacking model that integrates partial least squares regression (PLSR) and Support Vector Regression (SVR), achieving an impressive R² value of 0.963. The accuracy of models with the vegetation index as an input variable is also high. For example, Li et al. [45] utilized a BiLSTM model, which achieved an R² value of 0.624, while Zhang et al. [46] employed an MLP model, achieving an R² value of 0.638. In this study, it was found that the spectral reflectances with higher correlations with the SMC were red edge 750 and NIR, and the spectral indices with higher correlations were the DVI, NDVI, and RVI. The inversion accuracy with spectral reflectance as an input variable was better than when using the spectral indices. This may be due to the fact that band reflectance directly characterizes the spectral features of soil moisture, which can provide richer information, especially in bare soil or areas with low vegetation cover, whereas vegetation indices mainly reflect the state of vegetation under water stress, which has limitations such as providing less information, being indirect, and being susceptible to vegetation interference. It is worth noting that comparisons between spectral reflectance and spectral indices are still lacking in existing studies, and most of them focus on the topsoil layer; the applicability at different soil depths needs to be explored further in order to optimize the input variables and model selection for the multispectral inversion technique.

4.2. The Influence of Soil Background on the Accuracy of Soil Moisture Content Inversion

High-resolution multispectral images captured by drones may be significantly affected by soil background interference when inverting soil moisture content. This interference primarily arises from the mixed pixel effect, as well as the combined impact of various environmental factors such as meteorological conditions, the electromagnetic environment, and topography. Studies have shown that soil background interference has a dual nature. On the one hand, Zhang et al. [18] and Da et al. [19] found that removing the soil background significantly improves inversion accuracy, which may be because it stops bare soil spectra from interfering with the vegetation moisture signal; on the other hand, this study is consistent with the results of Wang et al. [20] and Zhang et al. [15], who showed that inversion accuracy is actually reduced after removing the soil background.

The decline in inversion accuracy may be associated with the following factors: (1) In the experimental area, the wind speed was relatively high, which partially flattened the alfalfa. This resulted in the loss of some effective canopy information when removing the soil background. Additionally, the shadow effect at midday further compromised the integrity of the spectral characteristics. (2) The temporal variability of crop moisture stress responses results in differences in the time windows for spectral signal acquisition. (3) At the plot level, the spectral information is a result of interactions between the vegetation indices and various environmental factors. Isolating vegetation information can distort the relationship between the spectral data and the measured values, ultimately leading to a decrease in inversion accuracy after the soil background is removed [47]. (4) The accuracy of soil moisture inversion models is influenced by various factors, including the crop type, environmental conditions, and sensor characteristics. Currently, the applicability of NDVI-based soil background removal methods in different vegetation systems requires systematic validation. When selecting the classification methods, this study utilized only the NDVI method for removing soil background noise. However, during NDVI classification, some areas with indistinct differentiation from plants, such as shadows, may be mistakenly classified as alfalfa plants. This could potentially decrease the inversion accuracy after soil background noise removal. Therefore, alternative methods, such as the RDVI and OSAVI methods, could also be employed to classify the alfalfa canopy and soil in the images. Research has demonstrated that the Optimization of Soil-Adjusted Vegetation Index (OSAVI) [48] and the ratio of the Transformed Chlorophyll Absorption in Reflectance Index (TCARI) [49] to the OSAVI are effective in mitigating the influence of soil background, outperforming the NDVI method in this regard. Furthermore, Zhang et al. [18] found that both the RDVI and OSAVI methods yielded superior classification results for plants and soil, resulting in improved inversion performance.

4.3. Uncertainty Analysis

This study found that using the KNN, RFR, RR, and XG-Boost models to invert the soil moisture content in the alfalfa field achieved good results; however, certain limitations persist. Firstly, limitations in sensor accuracy, atmospheric correction errors, and band calibration deviations can lead to reflectance inaccuracies, which, in turn, affect the precision of vegetation indices calculated from reflectance data. Additionally, environmental conditions can impact predictions of soil moisture content. For example, while drone remote sensing has been widely employed in agricultural monitoring, data acquisition is often restricted by weather conditions. Factors such as cloud cover thickness can affect the stability of spectral data, leading to fluctuations in band reflectance and shifts in image tones. High temperatures and drought conditions can induce changes in chlorophyll stress, subsequently altering canopy reflectance. Strong winds may also modify the canopy structure and affect the distribution of leaf angles, introducing noise into spectral data and ultimately reducing the accuracy of spectral indices. Additionally, rainfall before drone flights can change surface moisture levels, leading to fluctuations in reflectance values. When the observation direction of a drone’s sensor does not align with the direction of direct sunlight, the drone images often contain shadows that weaken canopy spectral information, thus compromising the accuracy of soil moisture inversion [50]. To address these issues, future research should focus on exploring the quantitative relationship between changes in weather conditions and the accuracy of model inversions. Additionally, synergistic application of satellite and drone remote sensing technologies should be investigated to overcome the limitations of a single data source. Moreover, this study primarily utilized data from multispectral sensors and did not integrate multimodal information such as thermal infrared and meteorological data. Multimodal information fusion technology, through collaborative inversion and modeling of multi-source remote sensing data, can significantly enhance high-precision dynamic monitoring of the soil moisture content. Zhang et al. [51] found that multimodal data fusion improved the accuracy of soil moisture content estimation. Regardless of the machine learning algorithm used, the optimal input variables were identified as a multimodal fusion of thermal and multispectral information. Zhu et al. [52] effectively fused 10 spectral bands with FVC features and applied ensemble learning to accurately predict the soil moisture contents at different root zone depths. Future research could integrate thermal imaging or synthetic aperture radar (SAR) data with knowledge-based models. By inputting thermal imaging and SAR data into these models, it would be possible to utilize the physical rules and empirical formulas within the models for soil moisture inversion. This approach not only has the potential to improve inversion accuracy but also provides more reasonable explanations and physical significance, thereby enhancing the credibility of the results. These improvements will help establish a more robust soil moisture monitoring system.

5. Conclusions

(1) In the model constructed using spectral reflectance, the XG-Boost model demonstrated strong performance both with and without removal of the soil background, achieving R² values of 0.812 and 0.803 on the validation set, respectively. In contrast, for the model based on spectral indices, the Random Forest Regression (RFR) model provided the best performance, regardless of whether the soil background was removed, with R² values of 0.632 and 0.628 on the validation set. (2) The inversion models using spectral reflectance as input variables demonstrated superior accuracy compared to those using spectral indices. The validation R² values for the XG-Boost, KNN, RFR, and RR models with spectral reflectance as the input variable generally ranged from 0.7 to 0.8, while the validation R² values for the models utilizing spectral indices typically ranged from 0.4 to 0.6. (3) After removing soil background noise, there was no significant improvement in the accuracy of alfalfa surface soil moisture inversion. The accuracy changes for the KNN, RFR, RR, and XG-Boost models were minimal. When using spectral reflectance as the input variable, the R² value for the highest-performing XG-Boost model only changed by 0.009. Similarly, for the top-performing RFR model using vegetation indices as input variables, the R² value for the validation set changed by just 0.004. The findings of this study provide theoretical and technical support for the prediction of alfalfa surface soil moisture. However, further investigation is needed to assess the applicability of these results to other crops and regions. Future research should use a wider variety of vegetation and conduct soil moisture inversion studies in multiple locations in order to enhance the accuracy and applicability of the models.