Evaluation of Cotton Defoliation Rate and Establishment of Spray Prescription Map Using Remote Sensing Imagery

Abstract

The site-specific management of cotton fields is necessary for evaluating the growth status of cotton and generating a defoliation prescription map. The traditional assessment method of pests and diseases is based on spot surveys and manual participation, which is time-consuming, labor-intensive, and lacks high-quality results. The RGB and multispectral images acquired by drones equipped with sensors provide the possibility to quickly and accurately obtain the overall data for a field. In this study, we obtained RGB and multispectral remote sensing images to calculate the spectral index of the target area. At the same time, ground survey data were obtained by tracking and investigating the defoliation rate of cotton after spraying. With the help of data analysis methods, such as univariate linear regression, multiple linear regression models, neural network models, etc., a cotton defoliation effect monitoring model based on UAV remote sensing images was constructed. The results show that the BP neural network based on the VARI, VDVI, RSI, NGRDI, NDVI index has an R² value of 0.945 and RMSE value of 0.006. The R² values of the multiple linear regression model are 0.844 based on the RSI and NGRDI indexes and RSI and VARI indexes. Additionally, based on the model, the cotton defoliation of the whole farmland was evaluated, and the spray prescription map of the UAV sprayer was obtained.

1. Introduction

The precision spraying of agricultural chemicals refers to a prescription map generated based on a navigation and positioning system and geographic information technology, or real-time sensor technology, to obtain different information about crops or diseases and insect pests in small farmland areas [1]. This information is converted into a spray prescription value that combines the different information within the decision support system [2]. The nozzle’s spray volume is adjusted through pressure or flow control to achieve differentiated spray operation [2]. Precise spraying can reduce the application volume of pesticides and improve the utilization rate of pesticides. Precision spraying is not a new concept, and there has been much research on targeted precision spraying in weed spot removal and orchards [3]. These spraying applications implement sensor-based real-time variables according to crop growth and plant characteristics. However, the control of diseases and insect pests needs to comprehensively consider the overall occurrence of farmland, which requires obtaining a spraying prescription map in advance. Campos et al. attempted to use remote sensing images to characterize the structure of the grapevine canopy in order to achieve variable spraying based on the grape canopy structure in the vineyard [4]. Their study establishes a linear relationship between ground survey and remote sensing data by combining manual ground calibration and remote sensing analysis. Furthermore, the canopy structure generated a variable spray prescription map for vineyards. Rudd et al. [5] incorporated cotton plant height and vegetation index NDVI into the main factors of variable spraying and expressed the difference in the cotton canopy by artificially setting factor weights to realize variable spraying of defoliants. How to obtain the spray prescription map based on the different information of crops is the focus of this study.

Cotton harvest aids are an essential step before mechanized harvesting [6,7]. Cotton harvest aids involve using chemical defoliants and ripening agents to interfere with cotton’s physiological and biochemical processes [8]. This accelerates the growth process and maturity of cotton, as well as making its leaves fall off early. In recent years, unmanned aerial vehicle (UAV) sprayers for cotton defoliation have been successfully used in cotton production areas in China [9,10]. However, the current UAV sprayers have not yet achieved precise spraying according to differences in types of farmland. Traditional spray methods rely on an applicator to set the same spray volume. This method may ignore the differences in crop growth in the field, resulting in the over-application or under-application of pesticides. Variable-rate spraying based on a prescription map helps to reduce the number of sprays, and the level of plant protection products applied [2]. Before the UAV sprayer second sprays, assessing the defoliation of cotton helps to provide data support for determining the mechanized harvest time and precise spraying.

The traditional evaluation of pesticide application is a very labor-intensive and time-consuming task. Statistical methods such as five-point or random sampling are manually performed in the test area [11,12]. This method has artificial statistical errors and lacks the high-quality results of the whole target area. Remote sensing technology makes it possible to rapidly detect large-scale farmland information [13]. Remote sensing technology receives electromagnetic radiation information from ground objects via satellites, airplanes, or ground sensing equipment, and then conducts ground object data analysis [14]. Because remote sensing is a non-contact technology, it will not affect the growth of crops. At the same time, remote sensing technology can quickly collect agricultural information on a wide range of fields [15,16,17]. Zhang et al. used satellite images to evaluate the spraying effects of fungicides and plant growth regulators on M-18B agricultural aircraft [18]. Their experiment analyzed the relationship between droplet deposition and vegetation index, and the results reveal that these two factors have a specific correlation. The study found that the vegetation indices NDVI and MSAVI calculated from satellite images can be used to evaluate the effect of agricultural aerial spraying on large-scale farmland. Ju et al. used the color, texture, and vegetation index of the target area to evaluate the spraying effect of herbicides [19]. The results show that the difference between the excess green index and the excess red index (EXG-EXR) could effectively detect the weeding effect. Ampatzidis et al. use a low-cost and rapid UAV-based phenotyping technique to evaluate citrus rootstock cultivars [20]. This automated and emerging technology can be used to assess individual trees (also known as phytotechnology) by analyzing plant-based phenotypic traits and reducing the number of personnel needed for manual data collection. The low-altitude remote sensing of UAV is easy to operate and low in cost compared with satellite remote sensing and ground remote sensing. This technique can obtain high-resolution images of the field and avoid the lack of macro-information caused by ground point collection.

Plant growth status and its reflection characteristics in a specific spectral band are correlated [21]. For multi-spectral images, because healthy vegetation appears as an absorption valley in the red region, the reflectance near the wavelength of 760 nm sharply increases, and the near-infrared region shows high reflection characteristics [22]. At the same time, soil, rock, water, etc., have no noticeable change characteristics in this wavelength region. Therefore, multispectral images, such as canopy coverage extraction [23], leaf area index calculation [24], pest monitoring [25] and chlorophyll estimation [26], are often used in applied research. In addition, UAV RGB images are often used in applications such as canopy coverage, crop height, and lodging rate calculation due to their clear images and high spatial resolution [27,28,29]. A series of physiological and biochemical changes occur during the cotton defoliation process. After spraying defoliants, ethylene and abscisic acid that promote shedding in leaves will increase, while the content of auxin that inhibits shedding will decrease [30]. First, the leaves of the plant change from tender green to gradually withered, and then the visible leaves fall off. At the same time, with the opening of cotton bolls, the color of the cotton field also changed from green to white. It may be feasible to evaluate cotton defoliation based on UAV multispectral and RGB images.

In this study, the remote sensing images of the whole process of cotton defoliation were tracked in the artificially intervened experimental fields. The main contents of the research include three parts. (1) Testing the feasibility of using UAV remote sensing images to evaluate the defoliation rate of cotton. (2) Predicting cotton defoliation based on vegetation index. (3) Evaluating the defoliation of farmland and constructing a spray prescription map suitable for UAV sprayer.

2. Materials and Methods

2.1. Field Plot

The experiment was carried out in Changji, Xinjiang (87°18′22″ E, 44°7′21″ N) in September 2019. The cotton test variety was Xinluzao 57, the sowing date was April 23, and the sowing density was 18,000 plants/ha. The area of the experimental field was 5.3 ha. The test pilot area has been a cotton planting region for many years, and the sowing intervals of the wide and narrow rows were 0.66 and 0.1 m, respectively. In the early stage of cotton farming, different irrigation management strategies were carried out in the test field for six plots. Moreover, the six experiment plots were sprayed with defoliant with different spray volumes before the experiment. As shown in Figure 1, the first spraying was carried out on the above plots, and the gradient range in the spray volume was 18–30 L/ha. The spraying equipment used was XAG P series Unmanned aerial spraying systems (XAG Co., Ltd., Guangzhou, China). The areas mentioned above were sprayed for the second time, and the spray volume was set as above. The defoliant applied were Thidiazuro Suspension concentrate 180 g/ha, “Banba” (alkyl ethyl sulfonate) 720 g/ha, ethephon 750 g/ha, and adjuvant “Beidaton” 150 g/ha. The initial values of cotton defoliation rate with differences were obtained through the above-mentioned experimental pretreatment in 6 plots of the test area.

2.2. Data Collection

2.2.1. Defoliation Rate Data

The cotton defoliation effect investigation (shown in Figure 2) was conducted under the Chinese national standard of DB65/T 3843.6-2015. Three pesticide application grids were selected for six plots to investigate the effect of defoliation. The survey points are fixed points, and each survey point counts for at least five cotton plants. The corresponding formula is shown below. The follow-up survey dates were 3 days, 5 days, 7 days, 10 days, 13 days, 15 days, and 17 days after spraying, a total of 7 times.

$$Defoliation rate\left(\%\right)=\frac{A-B}{A}100$$

(1)

where A is the total number of leaves of the plant before application, and B is the number of remaining leaves of the plant at investigation. The location information of the sampling points is obtained by the XAG UAV handheld surveying device (XAG Co., Ltd., Guangzhou, China). The horizontal positioning error of this device is ±1 cm.

2.2.2. Remote Sensing Image Data

The Phantom 4 RTK is equipped with a 20-megapixel CMOS sensor for capturing RGB images. The positioning module of the Phantom 4 RTK has real-time kinematic positioning (RTK), which can achieve a positioning accuracy of 1 cm ±1 ppm in the horizontal direction and 1.5 cm ±1 ppm in the vertical direction. In this study, the flight height of Phantom 4 RTK is 100 m, the previous overlap rate and side overlap rate are 80% and 70%, respectively, and the ground sampling distance (GSD) is 2.706 cm/pixel. The camera’s shutter speed is the key to ensuring the clarity of the image. The aperture F is set to the maximum value during the test to ensure sufficient light input. The gimbal mounted on the drone can ensure that the camera lens always faces vertically downward during the operation. Even if the pitch angle changes during the drone’s movement, the camera can maintain a fixed shooting angle.

The SenseFly eBee aircraft (senseFly, Lausanne, Switzerland) equipped with a Parrot Sequoia^® (Parrot, Paris, France) multispectral camera was used to collect multispectral remote sensing images. The Parrot Sequoia camera integrates a light sensor and a camera with four spectral bands, namely green (wavelength 550 nm), red (wavelength 660 nm), red edge (wavelength 735 nm), and near infrared (wavelength 790 nm). The flying height of the eBee SQ UAV is 80 m, and the previous overlap rate and side overlap rate were both 80%. The light sensor can pre-correct the image according to light conditions.

The acquisition date of the remote sensing image was the same as the date of the defoliation survey (Figure 3). Since the images acquired by different acquisition devices have different spatial and geographic locations, they need to be geometrically corrected. After the remote sensing images were stitched by the PIX4D software (Pix4D SA, Lausanne, Switzerland), the image registration tools in ENVI (Research Systems Inc., Boulder, CO, USA) were used to geo-correct images with different times and sensors. The high-precision image obtained by RTK was used as the reference image, and the polynomial model was used to determine the control points of the apparent features in the two images. With the increase in control points, the error gradually decreases. When the error is less than one pixel, the image registration process is completed. The region of interest (ROI) tool in ENVI was used to label the ground survey areas. Spectral index calculations were performed on all ROI regions using the Band Calculation tool in ENVI. Some vegetation indices that have been widely used in plant growth status prediction are selected in this study (

Table 1. The calculation formula of the remote sensing spectral index of the target area.

Vegetation Index	Formula	References
Visible-band Difference Vegetation Index (VDVI)	$\frac{2G-R-B}{2G+R+B}$	[31]
Normalized Green–Red Difference Index (NGRDI)	$\frac{G-R}{G+R}$	[32]
Visible Atmospherically Resistant Index (VARI)	$\frac{G-R}{G+R-B}$	[33]
Ratio Spectral Index (RSI)	$\frac{NIR}{R}$	[34]
Normalized Difference Vegetation Index (NDVI)	$\frac{NIR-R}{NIR+R}$	[35]
Red Edge Chlorophyll Index (CI)	$\frac{NIR}{RE}-1$	[36]
MERIS Terrestrial Chlorophyll Index (MTCI)	$\frac{NIR-RE}{RE+R}$	[37]

). RGB images were used to calculate visible-band difference vegetation index (VDVI), Normalized Green–Red Difference Index (NGRDI), and Visible Atmospherically Resistant Index (VARI). The Ratio Spectral Index (RSI), Normalized Difference Vegetation Index (NDVI), Red Edge Chlorophyll Index (CI), and MERIS Terrestrial Chlorophyll Index (MTCI) were calculated based on multispectral images.

2.3. Data Analysis

In order to screen the appropriate spectral index, the spectral index value of the target area was linearly fitted to the result of the defoliation rate. The high correlation index and the defoliation rate value were optimized to perform multiple linear regression models and neural network modeling and evaluate the accuracy of the established model.

2.3.1. Multiple Linear Regression

Multiple linear regression (MLR) is a statistical method that uses multiple explanatory variables to predict the outcome of a response variable [38]. Since a phenomenon is often associated with multiple factors, it is more effective to predict or estimate the dependent variable by the optimal combination of multiple independent variables, rather than using only one independent variable. The expression of the multiple linear regression model is:

$$y_i={\beta }_0+{\beta }_1{x}_{i1}+{\beta }_2{x}_{i2}+\dots +{\beta }_p{x}_{ip}+\epsilon$$

(2)

y is the dependent variable, x_i1, x_i2…, x_ip are independent variables, ${\beta }_0$ is the constant term, and, ε is the error term.

Establishing a multiple linear regression model requires that the independent variables and dependent variables meet the preset conditions. The linear relationship between the independent variables and the dependent variables, the independence between the residuals, and the residuals should obey the normal distribution. The residuals have a homogeneous variance, and the dependent variables are continuous variables, and independent variables are continuous variables or categorical variables. There is no multicollinearity between the independent variables, and the sample size should be more than 20 times the independent variable, etc. [39,40].

2.3.2. BP Neural Network

The neural network is an algorithmic mathematical model for distributed and parallel information processing by simulating the behavioral characteristics of animal neural networks [41]. It consists of a wide-ranging parallel interconnected network composed of adaptive neurons and achieves the purpose of processing information by adjusting the interconnection relationship between many internal nodes. A neuron is the basic unit of a neural network. The network accepts input signals from the weighted connections of n other neurons and outputs the signal through the activation function after adding a bias. Multiple neurons are connected in a hierarchical structure to form a neural network [42].

A neural network usually consists of an input layer, a hidden layer, and an output layer. Neural network training generally uses error propagation (back propagation, referred to as BP) for training. The BP algorithm is based on a gradient descent strategy to adjust the parameters in the direction of the negative gradient of the target [43]. The BP algorithm transfers the training set data to the input layer neurons, as well as the signals to the output layer layer-by-layer. It then calculates the error between the output layer signal and the true value, and then propagates the error back to the hidden layer neurons, and finally, according to the hidden layer. The error of the element adjusts the connection weight and offset. The entire iterative process loops until certain stopping conditions are reached. The goal of the algorithm is to minimize the cumulative error on the training set.

In this study, a Bayesian regularization training algorithm is used to train the neural network. The Bayesian regularization training algorithm introduces a correction function to the performance function on the basis of the neural network training function [35]. The network training error function expression is

$$E_D={\displaystyle {\displaystyle \sum }_{i=1}^n}{(t_i-t_x)}^2$$

(3)

where t_i is the actual output and x_i is the expected output. Both findings in this study were defoliation rates.

The mean of the sum of squares of all network weights is:

$$E_W=\frac{1}{m}{\displaystyle {\displaystyle \sum }_{i=1}^m}{w}_i{}^2$$

(4)

The performance function of the network is

$$F\left(w\right)=\alpha E_W+\beta E_D$$

(5)

where α and β are regularization coefficients, which affect the complexity and smoothness of the network, respectively. When α is too small, the network is overfitting. The β parameter affects the smoothness of the network. If β is too small, the network is underfitting.

The idea of Bayesian regularization algorithm is to set the weight parameters as random variables and determine the optimal weight function according to the probability density of the weights. The Bayesian regularization algorithm can effectively improve the overfitting problem during neural network training and reduce the error of fitting the curve. At the same time, it has a good generalization ability for data sets that are relatively small or noisy.

2.3.3. Model Accuracy Evaluation

In order to evaluate the accuracy of the model, the root mean square error RMSE and the coefficient of determination R² are used as the error analyses of the regression prediction algorithm. RMSE is the average error between the predicted results and the actual results of the defoliation rate. The predicted results come from the calculation of the model, and the actual results refer to the manual ground survey in Section 2.2.1. The lower the RMSE, the better the model. R² represents the square of the correlation coefficient between the observed actual result and the predicted value constructed by the model, and its value ranges from 0 to 1. The higher the value of R², the better the model.

$$RMSE=\sqrt{\frac{1}{m}{\displaystyle \sum }_{i=1}^m{\left(y_i-\widehat{y_i}\right)}^2}$$

(6)

$$R^2=1-\frac{{{\displaystyle \sum }}_i{\left(\widehat{y_i}-y_i\right)}^2}{{{\displaystyle \sum }}_i{\left(\overline{y_i}-y_i\right)}^2}$$

(7)

Since all the samples were manually collected and measured, the total number of samples was insufficient. K-fold cross-validation is a commonly used method for evaluating models, and it can solve the problem of overfitting [42]. A K-fold cross-validation was performed, and K was equal to 10 in this study. A total of 90% of the samples were used for modeling, and the process was repeated 10 times.

3. Results

3.1. Defoliation Rate Result

Figure 4 is a box diagram of the change of the defoliation rate in the test area within 17 days after spraying treatment. The numerical distribution of the leaf falling rate ranges from 0 to 1, where 0 represents the initial state and 1 represents a complete decrease. It can be seen from Figure 4 that the value of the defoliation rate gradually increases with the spraying time. The monitoring period covers the entire numerical range of the defoliation rate, which means that the data meet the modeling requirements. The defoliation process significantly accelerated after 18 September. Because the supplementary spray of defoliant and the increase in temperature on 18 September greatly improved the effect of defoliation. There are fewer sampling points in the range of 0.3 to 0.6 defoliation rate. The lack of data in this part may affect the accuracy of the model.

3.2. Data Modeling

3.2.1. Correlation Analysis

Based on the unary linear regression model, the results of multiple spectral indices and defoliation rates were fitted. It can be seen from Figure 5 and

Table 2. Model fitting effect parameters.

Vegetation Index	VARI	VDVI	RSI	NGRDI	NDVI	MTCI	CI
R2	0.679	0.720	0.765	0.686	0.706	0.167	0.231
RMSE	0.183	0.171	0.157	0.181	0.176	0.295	0.284

that the fitted R² range of VARI, VDVI, RSI, NGRDI, and NDVI ranges from 0.679 to 0.765,and RMSE value from 0.157 to 0.183. The results show that the above indexes have a good correlation with the results of defoliation rate, and the RSI index has the strongest correlation. The R² value of MTCI and CI index is lower than 0.2312, and the RMSE value is larger than other indices. Therefore, five indexes, including VARI, VDVI, RSI, NGRDI, and NDVI, were included in the scope of modeling parameters of the defoliation effect.

3.2.2. Multiple Linear Regression Modeling

Since the above spectral index has a strong linear correlation with the results of the defoliation rate, a multiple linear regression model was used for modeling. In this study, the five spectral indices and the results of the defoliation rate show a linear relationship, the independent variables and dependent variables are continuous variables, and the actual sample size is greater than 20 times for the five independent variables. Therefore, it is necessary to judge the residual results and multicollinearity further.

As shown in

Table 3. Linearity test results of five spectral index models.

Modeling	Sum of Square	Degree of Freedom	Mean Square	F	Significance
Regress	11.175	5	2.235	133.509	0.000
Residual	2.009	120	0.017	-	-
Total	13.183	125	-	-	-

, the F test was performed on the linear regression model of the five spectral indices and the defoliation rate, and the significance result p = 0.000, indicating that the multiple linear regression model is statistically significant. Except for constants, the t-test significance of each regression coefficient is more significant than 0.05, and the results are significantly different from the model significance, indicating that the regression model may have multicollinearity [36]. In addition, the tolerance of each coefficient is less than 0.1, and the variance expansion factor VIF is much greater than 10 (

Table 4. Test results of the coefficients of the 5 spectral index models.

Modeling	Unstandardized Coefficients		Standardized Coefficients			Collinearity Statistics
	β	Standard Error (SE)	β	t	Significance	Tolerance	VIF
(Constant)	1.278	0.161	-	7.955	0.000	-	-
VDVI	0.919	1.283	0.124	0.717	0.475	0.043	23.413
VARI	1.224	2.800	0.307	0.437	0.663	0.003	387.750
RSI	−0.314	0.172	−0.534	−1.828	0.070	0.015	67.170
NGRDI	−5.429	5.361	−0.810	−1.013	0.313	0.002	503.562
NDVI	−0.187	0.643	−0.080	−0.291	0.772	0.017	60.199

), indicating that the multiple linear regression model based on the above five spectral indices has a serious multicollinearity relationship.

In order to avoid the collinearity of the parameters and reduce the number of independent variables, the independent variable parameters were eliminated one by one through the backward elimination method and separately modeled and verified. The evaluation method involves adding or removing independent variables to determine if they significantly reduce the residual sum of squares of the model. After analyzing and testing the variables one by one, it was found that the test results of three sets of parameter combinations, including RSI and NGRDI index combination, VARI and RSI index combination, VDVI and VARI index combination, meet the multicollinearity test and residual result test.

Table 5. Comparison of model parameters of three sets of parameter combinations.

	RSI, NGRDI	VARI, RSI	VDVI, VARI
Prediction coefficient R2 a	0.844	0.844	0.741
Residual sum of squares (RSS)	2.018	2.027	3.368
F	340.268	338.502	179.466
Significance	0.000	0.000	0.000

^a R² is the adjusted coefficient of determination.

shows that the two combinations of RSI and NGRDI and VARI and RSI have the best coefficient of determination R² among the three sets of parameter combinations, and the results are both 0.844. At the same time, the residual sum of squares of the comparison for the two combinations is significantly smaller than the combination of VDVI and VARI, so either one of the two can be selected as a multiple linear regression model.

According to

Table 6. Test results of the coefficients of the spectral index RSI and NGRDI models.

Modeling	Unstandardized Coefficients		Standardized Coefficients			Collinearity Statistics		RMSE
	β	SE	β	t	Significance	Tolerance	VIF
(Constant)	1.317	0.058	-	22.829	0.000	-	-	0.127
RSI	−0.340	0.030	−0.578	11.360	0.000	0.482	2.077
NGRDI	−2.767	0.341	−0.413	−8.117	0.000	0.482	2.077

and

Table 7. Test results of the coefficients of the spectral index RSI and VARI models.

Modeling	Unstandardized Coefficients		Standardized Coefficients			Collinearity Statistics		RMSE
	β	SE	β	t	Significance	Tolerance	VIF
(Constant)	1.324	0.057	-	23.079	0.000	-	-	0.164
RSI	−0.343	0.030	−0.584	11.575	0.000	0.491	2.038
VARI	−1.624	0.201	−0.407	−8.065	0.000	0.491	2.038

, the multiple linear regression model based on RSI and NGRDI parameters (F = 340.268, p < 0.001) and RSI and VARI parameters (F = 338.502, p < 0.001) is statistically significant. The prediction results of the cotton defoliation rate can be explained by RSI and NGRDI spectral index or RSI and VARI spectral index with 84.4% probability. The RMSE values of Y1 and Y2 are 0.127 and 0.164, respectively. The respective calculation formulas are as follows:

$$Y_1=-0.34\ast RSI-2.767\ast NGRDI+1.317$$

(8)

$$Y_2=-0.343\ast RSI-1.524\ast VARI+1.324$$

(9)

3.2.3. Neural Network Modeling

The number of neurons in the hidden layer is determined by trial and error method. After the comparison of the coefficient of determination R², when the number of neurons is 7, the modeling effect is better. Therefore, the structure of the BP neural network is determined. As shown in Figure 6, the input data are NDVI, NGRDI, VARI, VDVI and RSI values, and the output data are the defoliation rate of cotton. In total, 70%, 15% and 15% of the data were used for the training set, test set, and validation set, respectively. The model with the best performance in the 10-fold cross validation set is shown in Figure 7. The R² value of the training set is 0.945, the RMSE value is 0.006, the R² value of the test set is 0.916, and the RMSE value is 0.009, indicating that the model has a good predictive effect.

3.3. Evaluation of Defoliation Rate

According to the results in Section 3.2, the prediction accuracy of the BP neural network is higher than that of the linear regression model. From the results of multiple linear regression Y1 and Y2, it can be seen that indices, such as RSI, NGRDI, and VARI, have a more significant impact on the prediction model. Due to a large amount of remote sensing image data and the purpose of easy calculation, the multiple linear regression model Y1 was used to predict the cotton defoliation rate of the experimental field in 2019. The trial field completed the first phase of defoliant spraying, and the prediction result of the cotton defoliation rate is shown in Figure 8. The color in the picture from blue to red indicates that the defoliation state of cotton shifts from low to high, and the defoliation rate shifts from 0.43 to 1. The red in the figure represents complete defoliation, and the blue shows incomplete defoliation. The figure shows the defoliation of cotton in different spatial locations. Compared with traditional manual surveys, this model improves work efficiency.

3.4. Establishment of Spray Prescription Map

Figure 8 shows a spatial information map of cotton defoliation, showing differences in growth between regions. It also cannot be directly used for cotton defoliant spraying. Advice from agronomists is needed for a variety of factors, from agricultural infographics to spray prescription maps. Combined with the spray width of the UAV sprayer, the accuracy of the positioning system, and the response speed of the spray system, the actual size of the prescription grid is set to 10 × 10 m. In addition, according to the suggestions of agricultural technicians, the current defoliation state is divided into three spray volume levels. Class A corresponds to a defoliation rate greater than 80%, Class B corresponds to a defoliation rate of 80–60%, and Class C corresponds to a rate of less than 60%. As a result, a recommended map for spraying prescription values of cotton defoliant was generated (Figure 9). The map contains high-precision geographic location information and the corresponding spray volume, and the UAV sprayer can implement different pesticide spray volumes in different spatial locations. This prescription map can be used to guide the second stage of cotton defoliant spraying.

4. Discussions

This study uses multi-temporal RGB and multi-spectral remote sensing images to construct a cotton defoliation monitoring model. The defoliation rate evaluation model can quickly and accurately evaluate the cotton fields, providing a new method for spraying effect evaluation and prescription maps. Compared with the study of Rudd et al. [5], this study did not consider the canopy structure of cotton, but directly listed the defoliation rate as the purpose of modeling, which made the spray prescription map more intuitive. The defoliation rate is a relative value, which is easily affected by the initial state. This study expanded the number of samples by increasing the number of plants at a single sampling site and increasing the time and space distribution of the survey. The target area was artificially intervened to divide the test plots with water and fertilizer irrigation at the beginning of the experiment. There were spatial differences in the initial state of the cotton, so with different spray volume treatments, the cotton defoliation progress of the experimental plots is also significantly different. Based on the above intervention treatment, it is possible to obtain experimental data covering the full-cycle defoliation state. The data source of this study is only one cotton growing season, and the defoliation of cotton is still affected by changes in outside temperature. In follow-up research, we will continue to extend the test period and include more cotton varieties to improve our model’s generalization ability.

This study shows a strong correlation between VARI, VDVI, RSI, NGRDI, NDVI, and the cotton defoliation rate, while the MTCI and CI index were poor. Both the MTCI and CI indexes include the calculation of red edge parameters. The red edge parameter represents unique data of vegetation remote sensing and has an apparent linear relationship with the canopy chlorophyll and nitrogen content and is often used to estimate LAI [44,45,46]. The red edge parameter is not sensitive in the process of cotton defoliation because the sensitivity of the red edge parameter to the crop largely depends on the fineness of the sensor’s spectral resolution [24]. The lower resolution of the multispectral camera in this study is responsible for this finding. Gao et al. discussed the correlation between red edge parameters, spectral index, and winter wheat LAI. Their study found that the correlation between the four red edge parameters of REP, Dr, SDr, Dr/Drmin, and LAI is generally lower than RSI, NDSI, NDVI, OSAVI, TVI, MSAVI, MTVI1, and MCARI2 [24]. For the correlation between the index and LAI, the research results also prove that the red edge parameter is not sensitive. The strong correlations shown by VARI, VDVI, RSI, NGRDI, NDVI, etc., are due to changes in the phenotypic structure of the cotton. Under the action of ethephon and thidiazuron, abscisic acid and ethylene are formed in the plant, which stimulates the senescence of leaves and leads to shedding.

The multiple linear regression model and the BP neural network model were used to establish the model. The results show that the accuracy of the BP neural network model is better than the multiple linear regression model. The regularization of the neural network can avoid the model’s multicollinearity and improve its fit [47]. Similar results were also obtained in the study of Xu et al. [42]. However, the multiple linear regression model in this study has two independent variables, which reduces the relative complexity of the model. In the application process, the neural network model can evaluate the effect of defoliation based on the evaluation accuracy. However, based on the perspective of quickly generating a prescription map, it may be more convenient to use a multivariate linear model.

It is essential to evaluate the defoliation rate and generate a prescription map for the site-specific management of cotton fields. Traditional spraying methods usually use constant operating parameters. This method can easily overlook the spatial variation characteristics of the farmland, which can neither reduce the application amount of plant protection products nor guarantee a better application effect. Precision agriculture is a management strategy that gathers, processes and analyzes temporal, spatial and individual data and combines them with other information to support management decisions according to the estimated variability for the improved resource use efficiency, productivity, quality, profitability and sustainability of agricultural production [48,49]. This study can evaluate the defoliation rate of cotton fields and further generate spraying prescription maps. Especially in the second defoliant spraying of UAV sprayer and other spray equipment, it provides decision support for realizing the precise management of cotton fields.

5. Conclusions

In this study, an evaluation model of cotton defoliation rate based on remote sensing images and a spraying prescription map of cotton defoliant suitable for UAV sprayer were proposed. The results show that VARI, VDVI, RSI, NGRDI, NDVI have a strong correlation with cotton defoliation changes, while the MTCI and CI index have a poor correlation. For the neural network model based on the VARI, VDVI, RSI, NGRDI, and NDVI indexes show, the R² value is 0.945, and the RMSE value is 0.006; for the multiple linear regression model based on the RSI and NGRDI indexes, Y_1 = −0.34 × RSI − 2.767 × NGRDI + 1.317; for the model based on the RSI and VARI indexes, Y_2 = −0.343 × RSI − 1.524 × VARI + 1.324; and the R² values of the two linear regression models are both 0.844.The modeling results show that the accuracy of the neural network test set is higher than that of MLR, but MLR uses fewer spectral feature indices. Additionally, based on the model, the cotton defoliation of the whole farmland was evaluated, and the spray prescription map of the UAV sprayer was obtained. It was determined that the UAV equipped with a camera can accurately obtain the cotton defoliation information of the entire farmland and quickly convert it into the prescription map required for spraying. This shows the feasibility of using UAV remote sensing images to generate the spray prescription map in cotton defoliation. This research can provide decision support for the follow-up UAV precise spraying.