The Study of Various Regression Models Establishment to Identify Farmland Soil Moisture Content at Different Depths Using Unmanned Aerial Vehicle Multispectral Data: A Case in North China Plain

Abstract

Soil moisture content is one of the most important soil indices for agriculture production. With the increasing food requirement and limited irrigation water sources, it is of great significance to accurately and quickly measure the soil moisture content for precision irrigation, especially in deficient agricultural areas, such as North China Plain. To achieve this goal, more attention was paid to the application of unmanned aerial vehicle multispectral reflectance technology. However, it was urgent to enhance the regression models between spectral data and soil realistic moisture content, and there were limited studies about the regression research on deep soil layers. Thus, the farmland of winter wheat–summer maize double cropping at Yongnian District, Hebei, North China, was selected as the study area. A six-band multispectral camera mounted on a low-altitude unmanned aerial vehicle (UAV) was used to obtain the field spectral reflectance with bands from 470~810 nm, and meanwhile, soil moisture content at different depths (10, 20, 30, 40, 50, and 60 cm) was measured after maize sowing. Unary linear regression (ULR), multivariate linear regression (MLR), ridge regression (RR), and an artificial neural network (ANN) were employed to establish regression models. The results demonstrated that the sensitive bands of spectral reflectance were 690 nm, 470 nm, and 810 nm. Those models all established significant regression at the depths of 0–20 cm and 40–60 cm, particularly at 10, 50, and 60 cm soil layers. However, for a depth of 20–40 cm, the prediction accuracy was generally lower. Among MLR, RR, and BP models, the MLR exhibited the highest identification accuracy, which was most recommended for the application. The findings of this study provide technical guidance and effective regression for the multispectral reflectance on the farmland of North China Plain, especially for deep soil layer moisture prediction.

1. Introduction

Soil moisture content is a basic soil parameter to indicate the dry or wet conditions in the soil, which plays an important role in meteorology, hydrology, and agriculture [1,2,3]. Sufficient water irrigation is the foundation for agricultural production, but global climate changes lead us to a hotter and more parched world. Drought and water scarcity have become two of the most serious problems in agriculture around the world [4]. Thus, how to achieve efficient agricultural water use to alleviate water scarcity and ensure food security is of great significance. In recent years, the application of new technologies for real-time monitoring of field water shortage and accurate irrigation management to improve the water use efficiency of crops has been a research hotspot [5,6]. To quickly and accurately obtain the soil moisture content of farmland is the key to achieving precise irrigation, and with the rapid development of remote sensing technology, unmanned aerial vehicle (UAV) multispectral imaging has become a reality with great application prospect because of its flexibility, easy control, real-time performance, and high image resolution characteristics [7,8]. Currently, the studies on the remote sensing monitoring of soil moisture content have focused primarily on the inversion of surface soil moisture content (depth ≤ 10 cm) [9,10]. Zhang et al. [11] monitored the soil moisture content at different depths of 1, 5, and 10 cm using a multispectral camera mounted on a drone. They established regression models to invert the soil moisture content using different spectral reflectance factors and found that the best monitoring depth was the surface layer of soil at 1 cm, followed by depths of 5 cm and then 10 cm. Wang et al. [12] used the multispectral camera method to monitor the soil moisture content at depths of 1, 3, 5, and 10 cm and indicated that under the backpropagation (BP) neural network, the determination coefficients of the soil moisture content inversion model for the four depths were 0.866, 0.800, 0.975, and 0.911, respectively. However, there were rare reports about the application of UAV multispectral images on the inversion of soil moisture at deep soil layers. It is well-known that the moisture of the soil surface is variable and that the soil stored deep in a profile is more similar to the farmland in moisture content and crop water absorption [13,14]. Thus, more research on the inversion of deep soil moisture content with remote sensing monitoring was urgently needed. Many models, including statistical and neural network models, are available for the remote sensing spectral inversion of soil moisture content [15,16]. Sunet et al. [17] used multi-source remote sensing to collaboratively invert the regional soil moisture content through MLR, least squares regression, ANN, and other regression models. The fitting value of the soil moisture content from the ANN was 0.892, with the coefficient of determination (R²) reaching 0.796. Chung et al. [18] estimated the regional soil moisture content in the Jinjiang River Basin in South Korea using radar images and an ANN. They constructed an optimal neural network structure based on a number of hidden layers, hidden neurons, and activation functions. Compared with the observed values, the correlation coefficient (r) was 0.85. Yin et al. [19] inverted the soil moisture content based on ground spectral measurements and active microwave remote sensing and found that the multistep regression model best simulated the soil moisture content (its R² being 0.482), followed by principal component regression. Partial least square regression exhibited the worst performance. The backpropagation (BP) network was the best predictive model; with an R² value of 0.792, its prediction accuracy and stability were better than those of the other regression models. It was suggested that those various models should be tried and compared for a better inversion prediction when establishing a UAV multispectral image application.

The North China Plain (NCP) was one of the largest crop production areas in China, with a long history of winter wheat–summer maize double cropping system. It produces approximately 45% of the national wheat and 30% of the national maize [13]. With the increasing food requirement, it is extremely important to maintain the NCP grain yield. In this region, the agricultural water demand is high, and irrigation water accounts for a large proportion of it. However, the serious water shortage has limited irrigation in this region and has become the greatest challenge to improving the water use efficiency for grain production [20]. For precision irrigation in this region, the UAV multispectral image technology has become more and more important. Thus, this study was conducted at the field in Yongnian District of Handan city, Hebei province, North China, using a low-altitude UAV equipped with a six-band multispectral camera to obtain spectral reflectance, simultaneously collecting soil moisture data at different soil depths (10, 20, 30, 40, 50, and 60 cm) to determine the sensitive soil moisture content bands. Different regression models, including ULR, MLR, RR, and ANN techniques, were used to establish the regression models for different soil layer depths to invert the soil moisture content. The findings of this study could enhance the soil moisture prediction models of multispectral reflectance, especially at the deep layer, and provide technical guidance for farmland precision irrigation in North China Plain.

2. Materials and Methods

2.1. Field Experiment

2.1.1. Experimental Area

The experiment was conducted in the Luoguan campus experimental field at the Hebei University of Engineering. The test area (114°35′ E, 36°46′ N) was located in the Yongnian District, Handan city, Hebei Province, North China, with a typical warm temperate semi-humid and semiarid climate. The annual average temperature is 13.5 °C, the frost-free period is 200 days, the annual sunshine duration is 2600 h, and the average annual rainfall is 548.9 mm. The average depth of groundwater in the region is 24.2 m, and the soil type in the study area is calcareous brown soil [21].

2.1.2. Experimental Arrangement

The field experiment was carried out on the farmland with winter wheat–summer maize double cropping cultivation. The fertilization and field management referred to the local standards. After the wheat harvest, the field was covered with straw and residual wheat plants and was then plowed using a rotary tiller. Maize (variety: Jifeng 223) was sown on 14 June 2020. After maize was sowed, we collected the multispectral data on 15 June 2020 and measured the soil moisture at different depths. We used the field moisture data to guide the first irrigation of the following winter wheat. Twelve experimental plots were established, each with an area of 10 m × 10 m. To minimize cross effects, each plot was spaced 1 m apart. The experimental setup is shown in Figure 1.

2.1.3. Remote Sensing Data Acquisition

This study focused primarily on collecting spectral information within a relatively small monitoring range. An MD4-1000 micro drone (Microdrones, Siegen, Germany) equipped with a Micro-MCA camera (Tetracam, Los Angeles, CA, USA) was used as the main data collection method. The Micro-MCA system is a small six-band multispectral sensor comprising six independent channels with a wavelength range of 450 to 900 nm, which was mostly used for soil moisture monitoring [22,23]. The camera has six lenses corresponding to six bands of wavelengths: 470 nm (blue), 550 nm (green), 660 nm (red), 690 nm (red edge), 710 nm (near-infrared), and 840 nm (near-infrared). A remote sensing image was obtained using a drone on 15 June 2020, with clear, windless weather and good visibility. This was the second day after sowing the maize in the experimental field. To ensure data accuracy and flight safety, the flight route was set using an 80% overlap rate in the flight direction and a 70% overlap rate in the lateral direction, with a flight height of 150 m and a flight speed of 6 m/s. The lens was facing downwards, and a standard whiteboard was arranged in the experimental area for radiation calibration before takeoff to obtain the spectral reflectance of the experimental area.

2.1.4. Ground Data Acquisition

The ground data were measured at the same time as the remote sensing image acquisition. Soil moisture content was measured using the traditional drying method [24]. Soil samples were collected using a soil auger at depths of 10, 20, 30, 40, 50, and 60 cm with approximately 200 g of soil per layer and then weighed in an aluminum box. The gravimetric soil water content was measured by taking the proportion of the loss of mass after oven-drying at 105 °C to the constant mass of dry soil. Each sampling had three replicates at different soil layer depths.

2.1.5. Data Processing

During the experiment, the RAW-format data files were separated into six single-band images using the PixelWrench 2.0 provided with the multispectral camera. The MgExi GPS Tool 1.0 was used to process the camera model of each single-band image in batches, and PIX4DMapper4.4.12 was used to stitch the images together based on the position information provided by the photo EXIF data, resulting in the final orthorectified image. The collected ASD calibration board data and drone remote sensing image data, which underwent geometric correction and stitching processing, were used to perform radiometric correction on the flight data. Finally, the ENVI 4.8 was used to fuse the six bands into a single remote sensing image containing six reflectance bands. The reflectance information was then extracted from the images, and data analyses were performed using Excel 2019 and IBM SPSS Statistics version 22.

2.2. Regression Models

2.2.1. Unary Linear Regression (ULR)

The ULR model establishes a linear regression equation for prediction based on the correlation between the independent variable and the dependent variable. The method of ULR establishment was described by Chen et al. [25] and Yang et al. [26]. For the 33 soil samples collected from different depths, two-thirds of the data were randomly selected as the training dataset, and the remaining one-third was used as the validation dataset.

2.2.2. Multivariate Linear Regression (MLR)

The MLR model indicates that a linear regression model involves a relationship between more than one independent and dependent variable. The MLR model can be defined as follows [27,28]:

$$y=x\beta +\epsilon .$$

(1)

$$y=\left[\begin{array}{l}{y}_1\\ y_2\\ \cdot \cdot \cdot \\ y_n\end{array}\right],x=\left[\begin{array}{l}1\cdot \cdot \cdot x_{1p}\\ 1\cdot \cdot \cdot x_{2p}\\ \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \\ 1\cdot \cdot \cdot x_{np}\end{array}\right],\beta =\left[\begin{array}{l}{\beta }_1\\ {\beta }_2\\ \cdot \cdot \cdot \\ {\beta }_p\end{array}\right],\epsilon =\left[\begin{array}{l}{\epsilon }_1\\ {\epsilon }_2\\ \cdot \cdot \cdot \\ {\epsilon }_n\end{array}\right],$$

(2)

where the observations vector of y is an n × 1 matrix, x denotes an n × p matrix of the regress or variables, β denotes an p × 1 vector of regression coefficients, and ε denotes an n × 1 vector of random errors.

2.2.3. Ridge Regression (RR)

According to the previous study [26], a linear ridge matrix regression model can be expressed as follows:

$$y=XW$$

(3)

where X denotes an n × p matrix, y denotes an n vector, and W denotes the vector of the regression coefficients.

By minimizing the following equation:

$${\left|\right|{\left(X^{T}X\right)}^{-1}-y\left|\right|}^2+{\left|\right|\alpha IW\left|\right|}^2$$

(4)

we can obtain the following result as the RR estimator:

$${\widehat{W}}_{ridge}={\left(X^{T}X-\alpha I\right)}^{-1}{X}^{T}y$$

(5)

Here, RR adds a constraint term to the loss function with respect to the parameters and is chosen to minimize the penalized sum of squares. Variance inflation factor (VIF) was used to correct the models, and when the value of VIF > 10, it was suggested that the collinearity is relatively high (more details in [24]).

2.2.4. BP Neural Network

Artificial neural networks (ANNs) are characterized by their parallel-distributed processing ability, high fault tolerance, intelligence, and self-learning ability. They combine information processing and storage with unique knowledge representation and intelligent adaptive learning abilities. The ANN is a complex network comprising a large number of simple elements connected together with high nonlinearity and the ability to perform complex logical operations and implement nonlinear relationships. Much research has shown that most actual work problems can be approximated using a BP network with one hidden layer and that a three-layer BP network can solve any complex correlation analysis problem. In general, a BP network comprises three layers, namely, the input, hidden, and output layers. Figure 2 shows the framework of a BP network [27].

The BP network changes the weights between the two neurons in any two layers and the threshold by minimizing the difference between the actual and expected inputs, expressed as follows:

$$Minimize(f(X\left)\right)={\displaystyle \sum _{i=1}^N{\displaystyle \sum _{j=1}^s\frac{{(O_j^i-{\widehat{O}}_j^i)}^2}{N}}}$$

(6)

where N denotes the number of samples, and X denotes the vector containing the weights and threshold. When the training process is complete, the weights and thresholds can be determined.

3. Results and Discussion

3.1. Field Experiment Unary Linear Regression (ULR)

To understand the relationship between the reflectance in different bands and the soil moisture content at different soil depths, the correlation coefficient between the spectral data of each band and the soil moisture content at different depths was calculated, and the results are detailed in

Table 1. Correlation coefficient between spectral reflectance and soil moisture content at various depths.

Depth (cm)	Band 1 (470 nm)	Band 2 (550 nm)	Band 3 (660 nm)	Band 4 (690 nm)	Band 5 (710 nm)	Band 6 (810 nm)
10	0.250	−0.026	−0.034	0.288	0.010	−0.064
20	0.154	0.085	0.037	0.102	−0.012	0.003
30	0.060	0.079	0.058	0.054	0.008	0.025
40	0.369 *	0.333	0.364 *	0.426 *	0.352 *	0.383 *
50	0.496 **	0.350 *	0.364 *	0.533 **	0.386 *	0.425 *
60	0.437 *	0.250	0.223	0.442 *	0.251	0.241

Note: **—the significant level of correlation at p ≤ 0.001; *—the significant level of correlation at p ≤ 0.01.

. It showed that there was no significant correlation coefficient between those six bands with the soil moisture at 10~30 cm depth, especially at 30 cm depth. The six bands at depths of 40 and 50 cm pass the significance test at p ≤ 0.01, indicating a significant correlation. Additionally, at a depth of 60 cm, Bands 1 and 4 pass the significance test, indicating a significant correlation.

To visually analyze the data in Table 1, we plotted the correlation coefficients between the soil moisture content and spectral reflectance at different soil depths for each band, as shown in Figure 3. From Figure 3, it is evident that the correlation between the spectral reflectance and soil moisture content at soil depths of 50 and 40 cm is the most significant, with the highest correlation coefficients, followed by 60 cm. The sensitive bands for different soil depths are essentially the same. The most sensitive band is Band 4, followed by Bands 1 and 6, Bands 3 and 5, and finally Band 2. On the basis of the above analysis, the optimal monitoring depth for the soil moisture content for each band is 40–50 cm, with Band 4 being the best monitoring band, followed by Bands 1 and 6.

The ULR model was established using soil moisture data for depths of 40, 50, and 60 cm in the training dataset and their corresponding sensitive bands (Band 4). The model was then analyzed using Root Mean Squared Error (RMSE) and F-tests.

Table 2. The ULR model for sensitive bands.

Depths	Regression Equations	RMSE	F	p
60	y = 0.222b − 13.343	2.864	7.521	0.010
50	y = 0.239b − 14.257	2.419	12.327	0.001
40	y = 0.164b − 4.939	2.216	6.882	0.013

Note: y is soil water content; b is the reflectance of the band at 690 nm.

presents the results of the ULR model.

The results show that the ULR model for a depth of 40 cm is the best, with RMSE = 2.216 and F = 6.882 at a significance level of p = 0.013. The ULR model also exhibits a good fit for depths of 50 and 60 cm, although it is slightly lower than that at 50 cm. All models can be considered statistically significant at p < 0.05.

3.2. Multivariate Linear Regression (MLR)

As shown in

Table 3. The MLR model for all depths.

Depths	Regression Equations	RMSE	F	P
60	y = 0.511b1 + 1.319b2 − 1.868b3 + 0.7904b4 − 1.022b5 + 1.131b6 + 86.511	0.006	19.756	0.006
50	y = 0.596b1 + 0.262b2 − 0.619b3 + 0.613b4 − 1.170b5 + 1.032b6 + 52.157	0.006	6.615	0.04
40	y = 0.237b1 − 0.481b2 + 0.555b3 + 0.314b4 − 0.739b5 + 0.314b6 + 5.667	0.009	0.460	0.811
30	y = −0.313b1 + 0.998b2 − 0.895b3 + 0.337b4 − 0.437b5 + 0.436b6 + 55.811	0.017	0.450	0.818
20	y = 2.147b1 − 2.624b2 + 3.416b3 + 0.296b4 − 2.604b5 − 0.111b6 − 17.376	0.008	2.148	0.24
10	y = 2.375b1 − 1.744b2 + 1.095b3 + 0.006b4 − 1.192b5 + 0.168b6 + 36.507	0.007	30.036	0.003
0–20	y = 2.261b1 − 2.184b2 + 2.256b3 + 0.151b4 − 1.898b5 + 0.028b6 + 9.566	0.039	6.44	0.046
20–40	y = 0.69b1 − 0.702b2 + 1.025b3 + 0.316b4 − 1.260b5 + 0.213b6 + 14.701	0.191	2.643	0.183
40–60	y = 0.448b1 + 0.367b2 − 0.644b3 + 0.572b4 − 0.977b5 + 0.826b6 + 48.112	0.053	4.208	0.093

Note: y is soil water content; b₁, b₂, b₃, b₄, b₅, and b₆ are the reflectance of the six bands at 470 nm, 550 nm, 660 nm, 690 nm, 710 nm, and 840 nm, respectively.

, the MLR model is convenient and quick for practical applications and has significance for large-scale rapid monitoring of soil moisture in study sites. This study attempted to establish an MLR model for all soil depths to obtain a more accurate and suitable monitoring model.

It was evident that RMSE first increased and then decreased with the increasing soil depth, and a turning point appeared at 30 cm, which is the maximum value. The minimum RMSE was evident at soil depths of 10, 50, and 60 cm.

In summary, the MLR model for soil depths of 10, 50, and 60 cm exhibited the highest degree of fit with the six bands. The model for the soil depth of 20 and 40 cm with six bands had a slightly lower degree of fit. The fit degree for soil depths of 30 cm with the six bands was poor and statistically insignificant. These results may be due to the following reasons: The dryland soil was divided into four layers—the tillage, plow bottom, topsoil, and subsoil layers. The experiment was conducted on bare soil on the second day after sowing, and the field was tilled using a rotary tiller to a depth of approximately 0–15 cm, which was the soil layer with the greatest moisture change. Consequently, the surface soil moisture had the highest degree of fit with the MLR model for the six bands. The plow bottom layer at a depth of 20 to 30 cm exhibited poor ventilation and water permeability, which were poorly correlated with the reflectance of each band. Therefore, the degree of fit of the MLR model for soil moisture in this layer with the six bands was poor.

The topsoil layer at a depth of 30 to 50 cm is a water- and nutrient-retention layer that was not correlated with the reflectance of each band and lacked statistical significance. The subsoil below 50 cm is a natural soil layer. One watering session was conducted four days before sowing, and after four days of evaporation and seepage, the soil moisture of the surface and subsoil layers tended to be consistent. Consequently, the soil moisture in this layer exhibited a high degree of fit with the MLR model for the six bands.

3.3. Ridge Regression (RR)

The RR model is a biased estimation regression method that specializes in collinear data analysis. This sacrifices partial information and reduces the accuracy of obtaining more realistic and reliable regression coefficients. Collinearity analysis showed that the reflectance of different bands exhibited high collinearity issues in the inversion of soil moisture in different soil layers. Although the MLR model exhibited a high degree of fit, to obtain a more realistic inversion model, this study attempted to develop an RR model, and the results are shown in

Table 4. The RR model for different bands.

Soil Depth	Regression Equations	RMSE	p
10	y = 0.0075b1 − 0.0021b2 − 0.0021b3 + 0.0032b4 − 0.0031b5 − 0.0008b6 + 0.3075	0.014	0.015
20	y = 0.0054b1 + 0.0011b2 + 0.0002b3 + 0.0022b4 − 0.0044b5 − 0.0022b6 + 0.2247	0.079	0.253
30	y = 0.0009b1 + 0.0007b2 − 0.0001b3 − 0.0001b4 − 0.0012b5 + 0.0001b6 + 0.2341	0.119	0.423
40	y = 0.001b1 − 0.0004b2 − 0.0001b3 + 0.0001b4 − 0.0009b5 + 0.0007b6 + 0.0828	0.033	0.088
50	y = 0.0037b1 − 0.0008b2 − 0.0011b3 + 0.0020b4 − 0.0020b5 + 0.0017b6 + 0.0601	0.019	0.05
60	y = 0.0062b1 − 0.0005b2 − 0.0020b3 + 0.0026b4 − 0.0030b5 + 0.0012b6 + 0.1206	0.016	0.04
0–20	y = 0074b1 − 0.0005b2 − 0.0007b3 + 0.0031b4 − 0.0052b5 − 0.0012b6 + 0.3003	0.038	0.198
20–40	y = 0.0031b1 + 0.0006b2 + 0.0001b3 + 0.0014b4 − 0.0035b5 + 0.0000b6 + 0.2147	0.111	0.385
40–60	y = 0.0043b1 − 0.0007b2 − 0.0013b3 + 0.0022b4 − 0.0029b5 + 0.0020b6 + 0.1273	0.047	0.098

Note: y is soil water content; b₁, b₂, b₃, b₄, b₅, and b₆ are the reflectance of the six bands at 470 nm, 550 nm, 660 nm, 690 nm, 710 nm, and 840 nm, respectively.

.

From the results shown in Table 4, it is evident that the RMSE of the regression equations for the soil layer at different depths first increases and then decreases with increasing soil depth, reaching a maximum at 30 cm. Additionally, the RR model exhibits the highest RMSE at soil depths of 10, 50, and 60 cm, all of which are smaller than 0.02. The RMSE is 0.014 at a soil depth of 10 cm and exhibits the smallest degree. At soil depths of 50 and 60 cm, the RMSE of the RR model is approximately 0.02, indicating a slightly lower degree. At a soil depth of 40 cm, the RMSE of the RR model is 0.033, indicating a even lower degree of fit. At soil depths of 20 cm, the RMSE of the RR model is 0.079, representing a considerably poorer degree of fit.

From the above analysis, it is evident that the RR model’s degree of fit for soil moisture at different depths was consistent with that of the MLR model. Although its fitting accuracy was slightly inferior to that of the MLR model, the RR model was closer to the actual situation (owing to its eliminating collinearity) and proved to be superior to the MLR model.

3.4. BP Neural Network

To investigate the soil moisture content, this study designed a BP model with reflectance at different bands as input factors and different soil moisture content as output factors. As shown in Figure 4, it is evident that there is an enormous difference between the results from the BP model and the actual values for the soil layers at different depths. The RMSE of the model determined at different depths first increases before decreasing with increasing soil layer depth. A turning point appears at 20 cm, which is the maximum value. The RMSE is relatively small at soil depths of 10, 50, and 60 cm, indicating good fitting performance. Moreover, the RMSE coefficients at 10 and 60 cm are smaller than 0.005, and the degree of fit is the highest. At soil depths of 30 and 40 cm, RMSE is approximately 0.010, with the second-best degree of fit. At a soil depth of 20 cm, the degree of fit is the lowest.

From the above analysis, it could be concluded that the changes in the degree of fit of the BP model for soil moisture content at different depths were consistent with those of the MLR and RR models. However, the RMSE of the BP model was smaller, and the fitting effect was better than that of the RR model.

Scatter plots of the predicted and actual soil moisture content values at different depths were constructed for the MLR, RR, and BP models, and the fitting curves were compared. The results are shown in Figure 5, where the points of different shapes represent the predicted values of different models, and the lines of different colors represent the regression curves of different models.

As shown in Figure 5, the MLR, RR, and BP models exhibit high prediction accuracy at each depth, from 0 to 60 cm. The MLR model is the most accurate, followed by the BP model, and the RR model is the least accurate. This is consistent with the results of a previous study [19]. With the increase in soil depths, the RMSE first increased and then decreased. The lowest inflection point occurs at a depth of 20 to 30 cm before rebounding. Additionally, the lowest RMSE for the three models is at the 10, 50, and 60 cm soil layer depths, indicating that the regression effect is the best. At a depth of 10 cm, the RMSE values of the three models are all smaller, indicating that the models exhibit a good prediction ability at this depth. The RMSE of the three models at soil layer depths of 20, 30, and 40 cm are higher than other depths, indicating that the predictive ability of the three models is poor at these depths.

In conclusion, it is feasible to use UAV multispectral remote sensing technology for rapid identification of soil moisture content. The MLR, RR, and BP models exhibited high accuracy in fitting and predicting soil moisture content at depths of 10, 50, and 60 cm and offered good quantitative prediction capabilities. For the three depths, the RMSE of the MLR model was 0.007, 0.006, and 0.006, respectively, and the RMSE of the BP model was 0.004, 0.007, and 0.005, respectively. For depths of 0 to 20 cm and 40 to 60 cm, the three models exhibited good predictive capabilities. When predicting the soil moisture content at a depth of 20 to 40 cm, the four methods generally performed poorly, among which the BP model performed the best.

The fitting effect of each model at a depth of 20 to 40 cm was relatively low, which could be due to the influence of cultivation or the clay interlayer. Dry land soil can generally be divided into four layers, namely, the plow, plow sole, subsoil, and bottom soil. The most common cultivation method in the study area was rotary tillage at a depth of approximately 0 to 15 cm.

The experiment was conducted on the second day after sowing when the soil was bare. The soil was loosened and uniform after rotary tillage. At this depth, MLR at all six bands exhibited good fitting effects. The depth of the plow sole layer was approximately 20–30 cm, which could not be tilled by rotary tillage, forming a slowly developing interlayer with poor air permeability and water infiltration. This resulted in a poor correlation with the reflectivity of the six bands and, consequently, a lower fit for the MLR model. The depth of the subsoil layer was approximately 30–50 cm, which is a water and fertilizer conservation layer, and exhibited no correlation with the reflectivity of the six bands, making it statistically insignificant. The bottom soil layer, which is a natural soil layer, was below 50 cm in depth. Water was irrigated once four days before sowing, and after four days of evaporation and leakage, the soil moisture in the surface layer was consistent with that in the bottom layer. Therefore, for the MLR model, the soil moisture in this layer exhibited a good fit to the six bands.

Soil moisture can have a major impact on the identification results, and previous studies [29,30] have indicated that variations in soil moisture under wet conditions can increase the fitting error. Four days before the data collection day, the field was irrigated, the water gradually infiltrating the soil layer at a depth of 0 to 15 cm, whereas the soil layer at a depth of 20 to 40 cm was not tilled, and an interlayer was formed. The interlayer accumulated water and exhibited slow infiltration, explaining why the identification accuracy at depths of 20 to 40 cm was low.

From the above analysis, the MLR model proved to be the best, followed by the BP model, whereas the RR model exhibited the worst identification performance. However, for the MLR model, the serious collinearity of the spectral reflectance in each band resulted in overfitting. Therefore, in practical applications, the BP model should be prioritized to identify soil moisture content.

4. Conclusions

In summary, the sensitivity of the soil moisture content to spectral bands at different depths was essentially the same; the most sensitive band was Band 4 (690 nm), followed by Band 1 (470 nm) and Band 6 (810 nm). Band 3 (660 nm) and Band 5 (710 nm) were not recommended.

Comparing the performance of different models at various depths from 0 to 60 cm, the MLR model exhibited the highest identification accuracy, followed by the BP model, and the RR model exhibited the lowest identification accuracy. However, because of the serious collinearity among the reflectance of various spectral bands, there were overfitting problems for the MLR model. Consequently, the BP model is preferred in practical applications for identifying soil moisture content.

UAV multispectral remote sensing has been shown to be feasible for the rapid identification of soil moisture content. The three models, including the MLR, RR, and BP, exhibited high accuracy in predicting soil moisture content at depths of 10, 50, and 60 cm with good quantitative predictive capabilities, which could be used to monitor field soil moisture content. For depths of 0 to 20 cm and 40 to 60 cm, the three models exhibited good quantitative predictive capabilities; however, for predicting the soil moisture content at depths of 20 to 40 cm, MLR performed best during those models.

This study provides technical and theoretical support for an optimal monitoring depth and identification model of UAV multispectral remote sensing for soil moisture content, which is of great importance for water-saving and precision irrigation at the farmland in North China Plain.