Prediction method of cotton plant growth status during topping period based on UAV multispectral remote sensing

Abstract

The present invention discloses a method for predicting the plant growth state during the cotton topping period based on multispectral remote sensing of unmanned aerial vehicles, which comprises the following steps: S1, using a multispectral camera to collect multispectral information of the cotton canopy before and after the topping operation; collecting crop indicators related to the determination of the topping time strategy; S2, constructing a multispectral regression analysis model of the cotton canopy multispectral information-crop indicators; S3, using the multispectral information of a specific band combination collected from the cotton field to be predicted as input data, and obtaining the crop indicators through the multispectral regression analysis model. The method of the present application can realize the intelligent and rapid acquisition of cotton plant height, flower buds and fruit branches, and provide data support for the determination of cotton topping time.

Description

Technical Field

[0001] The present invention belongs to the field of smart agriculture, and in particular to a method for predicting the growth state of cotton plants during the topping period based on multispectral remote sensing of unmanned aerial vehicles. Background Art

[0002] Cotton is an important economic crop in my country. Xinjiang is an important cotton producing area in China. Xinjiang's cotton production is highly mechanized. Rapid and accurate monitoring of cotton growth status, precise management of cotton fields, and reasonable arrangement of production plans are of great significance to improving Xinjiang's cotton production efficiency and increasing cotton yields.

[0003] The cotton bud stage is one of the important growth periods of cotton. Cotton topping is an important task in the cotton bud stage. Scientific topping and fertilization are conducive to promoting the rational transformation of photosynthetic products and improving cotton yield and quality. The "Guiding Opinions on Cotton Mechanized Production Technology in the Northwest Inland Cotton Area" issued by the Ministry of Agriculture and Rural Affairs and related cotton planting agronomy requirements on cotton topping time: Determine the appropriate topping time according to factors such as cotton growth, plant height and number of fruit branches, topping early in time, and promote early maturity. Cotton topping has a strong timeliness, and the choice of operation time has a great impact on cotton growth. The determination of the operation window period in the operation strategy and path planning of drone or mechanical topping depends on the monitoring of crop growth conditions in the field. Therefore, rapid and accurate monitoring of cotton during the topping operation period can provide a crop data basis for scientifically determining the topping time and the strategic arrangement of topping operations. At present, the determination of cotton topping period mainly relies on manual field observation of crops, weather and soil conditions, and judging the topping time based on experience. The information acquisition efficiency is low and lacks interpretability. The main indicators for judging the topping period of cotton include: plant height, number of fruit branches and number of flower buds, etc. There are studies on the rapid acquisition of cotton plant height, but there are few studies on the rapid acquisition of information such as the number of fruit branches and flower buds.

[0004] Low-altitude UAV remote sensing has the advantages of high mobility, fast and flexible, low cost, and easy operation. It has important application prospects in precision agriculture and agricultural production management. At present, research results have shown that the use of UAVs equipped with color, multispectral, hyperspectral and other cameras to collect cotton canopy remote sensing information and establish regression models can monitor the plant height, flower buds and fruit branches of cotton. Commonly used regression models include multivariate linear regression, support vector machine, neural network, adaptive competitive weighting, decision tree, etc. At present, the research mainly focuses on the prediction of physiological indicators of cotton in a single or multiple growth periods. The data collection interval is long and the growth status of plants varies greatly. Therefore, the feature extraction ability of the prediction model is relatively low. Ensemble learning can improve the feature extraction ability of the model by combining multiple simple models. It can be applied to multispectral information processing to improve the model prediction effect.

[0005] In summary, the current determination of cotton topping period mainly relies on manual observation of crop growth status, which lacks interpretability and has low information collection efficiency. Using drones to collect multispectral information of crop growth can quickly and efficiently obtain plant growth status, but there is currently a lack of multispectral information processing models for continuous monitoring of plant growth status. Summary of the invention

[0006] This application aims to solve the problems in the background technology, takes cotton plants in the topping period as the research object, uses drones equipped with multispectral remote sensing cameras to collect cotton canopy remote sensing information every day, and measures plant indicators that affect the time and strategy of cotton topping. A regression model is established to analyze the monitoring status of plant growth status by each band and band combination information, and an integrated learning method is used to establish a regression model to enhance the regression effect of the model, providing a method for quickly acquiring and processing crop data for determining the time and strategy of cotton topping.

[0007] Technical solution:

[0008] A method for predicting the growth status of cotton plants during topping period based on multispectral remote sensing by unmanned aerial vehicles comprises the following steps:

[0009] S1. Use a multispectral camera to collect multispectral information of the cotton canopy before and after topping operation; collect crop indicators related to the determination of topping time strategy;

[0010] S2. Construct a multispectral regression analysis model of cotton canopy multispectral information-crop index;

[0011] S3. Using the multispectral information of the specific band combination collected from the cotton field to be predicted as input data, the crop index is obtained through the multispectral regression analysis model.

[0012] In S1, a drone equipped with a multispectral camera is used to obtain multispectral information of the cotton canopy.

[0013] In S1, a standard plate of fixed size is placed in the cotton field to be predicted, and the spectrum of the standard correction area is obtained as the radiation correction data of the remote sensing data. The correction formula is:

[0014]

[0015] Where I represents the average spectral value of the region of interest in a certain band, W represents the spectral mean of the standard white plate correction area of ​​the band on that day, B represents the pixel mean of the band when the lens is covered on that day, and CI is the spectral reflectance of the band on that day after radiation correction.

[0016] In S1, the crop indicators related to the topping time strategy include: cotton plant height, number of fruit branches, and number of flower buds.

[0017] In S2, the multispectral regression analysis model was constructed using the Adaboost+decision tree combination method.

[0018] The algorithm steps of Adaboost+decision tree are:

[0019] Input: training data set T = {(x1,y1),(x2,y2)…,(x n ,y n )},in ,y i ∈y, X is the instance space, y is the label set;

[0020] Output: Final classifier: G(x);

[0021] (1) Initialize the weight distribution of training data:

[0022]

[0023] Among them, w 1i is the weight of the initialization data, and N is the number of training data.

[0024] (2) For m = 1, 2, ..., M

[0025] (a) Using a weighted distribution D m The training data set is used to learn the basic regression model:

[0026] G m (x):X→y

[0027] G m is the final classifier;

[0028] (b) Calculate G m (x) Error rate e on the training data set m :

[0029]

[0030] w mi represents the weight of the i-th instance in the m-th round, Description m (x) The regression error rate on the weighted training dataset is G m The sum of the difference weights between the predicted value and the actual value of (x), from which we can see the weight distribution D of the data m With the basic classifier G m (x) the relationship between the classification error rate;

[0031] (c) Calculate G m Coefficient distribution of (x)α m :

[0032]

[0033] (d) Update the weight distribution of the training data set:

[0034] D m+1 ={w m+1,1 ,…,w m+1,i ,…,w m+1,N}

[0035]

[0036] Here Z m is the normalization factor:

[0037]

[0038] It makes D m+1 becomes a probability distribution;

[0039] (3) Construct a linear combination of basic regression models:

[0040]

[0041] The final regression model is obtained:

[0042]

[0043] In S3, the specific wavelength band combination is 550nm, 730nm, and 790nm.

[0044] Beneficial effects of the present invention

[0045] The method of the present application can realize the intelligent and rapid acquisition of cotton plant height, flower buds and number of fruit branches, and provide data support for determining the cotton topping time.

[0046] The above data can provide crop data support for the operation strategy and operation path decision-making of intelligent machinery.

[0047] 730nm and 790nm have good fitting effects on cotton plant height, fruit branches and bud numbers, 550nm has the second best fitting effect, which supplements the red light band information, and 660nm has the worst fitting effect. The 550+730+790nm band combination has a better fitting effect than other band combinations and single bands.

[0048] The Adaboost+decision tree method was used to analyze the multispectral data of the 550+730+790nm band combination. The model test results showed that the correlation coefficient between the canopy multispectral remote sensing information and the cotton plant height was 0.85, the predicted mean variance was 0.39cm, the correlation coefficient with the number of fruit branches was 0.84, the predicted mean variance was 0.49, the correlation coefficient with the number of flower buds was 0.84, the predicted mean error was 0.54, and the prediction result had an accuracy rate of 94.03% for the topping time. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a 660nm spectrum image collected by the unmanned aircraft in the embodiment

[0050] Figure 2a Schematic diagram of the average reflectance of samples collected in the 550nm band during the 12-day data collection period

[0051] Figure 2b Schematic diagram of the average reflectance of samples collected at 660nm during the 12-day data collection period

[0052] Figure 2c Schematic diagram of the average reflectance of samples collected in the 730nm band during the 12-day data collection period

[0053] Figure 2d Schematic diagram of the average reflectance of samples collected at 790nm during the 12-day data collection period

[0054] Figure 3a Schematic diagram of crop height measurement during 12 days of data collection

[0055] Figure 3b Schematic diagram of the measurement of the number of fruit branches of crops within 12 days of data collection

[0056] Figure 3c Schematic diagram of the measurement of crop flower bud numbers during the 12-day data collection period

[0057] Figure 4a Schematic diagram of prediction of plant height index during cotton topping period based on Adaboost+decision tree

[0058] Figure 4b Schematic diagram of the prediction of the number of fruit branches of cotton plants during topping period based on Adaboost + decision tree

[0059] Figure 4c Schematic diagram of the prediction of the number of flower buds in cotton plants during topping period based on Adaboost + decision tree DETAILED DESCRIPTION

[0060] The present invention will be further described below in conjunction with embodiments, but the protection scope of the present invention is not limited thereto:

[0061] 1 Materials and methods

[0062] 1.1 Overview of the experimental area and experimental materials

[0063] The experimental cotton variety in the embodiment is Xinlu Zao 68. The data collection time is from July 1, 2021 to July 12, 2021, and the multispectral remote sensing information of the cotton canopy and the crop growth indicators affecting the time and strategy of cotton topping are continuously collected for 12 days before and after topping. The experimental field was toppled on July 8. The data collection location is the location of the Eighth Division Corps in Shihezi, Xinjiang. The location of the experimental field is 84° east longitude and 43° north latitude, with an area of ​​6.75 acres. Shihezi City has a typical temperate continental climate with an average temperature of 25.1-26.1°C, an annual precipitation of 125.0-207.7mm, and long sunshine time in summer. Cotton is an important economic crop in Shihezi. The cotton planting area accounts for more than 85% of the crop planting area, and the scale, intensiveness, and large-scale mechanization are the first among the three major cotton areas in China. It is an important base for the development of intelligent cotton production.

[0064] 1.2 Data collection and preprocessing

[0065] 1.2.1 Multispectral Data Collection

[0066] The experiment uses the XAG Jixia drone, which has a body weight of 10kg and a maximum load of 6kg. Use a smartphone to log in to the XAG control system and check the target plot. The drone flies according to the route automatically planned by the system. The drone flies at an altitude of 7m and a speed of 3m / s. The data collection time is 2 pm every day. The multispectral camera covers four bands, with wavelengths of 550, 660, 730, and 790nm respectively. The multispectral camera has 20 million pixels, an image resolution of 3863*3648pixel, a spectral resolution of 10nm, and a camera mass of 0.85kg, which is suitable for operation on a drone flight platform. During the experiment, a 50*50cm standard whiteboard was placed on the experimental plot as radiation correction data for remote sensing data. Unmanned aircraft image remote sensing experiment as shown in the figure. Figure 1 shown.

[0067] 1.2.2 Multispectral Data Preprocessing

[0068] After the multispectral remote sensing data of cotton canopy is acquired, it needs to be preprocessed. The spatial resolution of the remote sensing image is 0.02m*0.02m. The remote sensing data preprocessing mainly includes two steps. (1) Selection of the region of interest. For each spectral band of each image, a 50*50pixel rectangular box is used to manually select the area of ​​cotton in the image as the region of interest, and the average value in the region of interest is calculated as the spectral value of the image in this band. (2) Radiation correction of multispectral remote sensing images. A 30*30pixel rectangular box is used to select the position of the standard white board as the white correction area, and the spectral values ​​of the white correction area are averaged. The multispectral remote sensing data is radiation corrected using formula (1). Where I represents the average spectral value of the region of interest in a certain band, W represents the spectral mean of the standard white correction area of ​​the band on that day, B represents the pixel mean of the band when the lens is covered on that day, and CI is the spectral reflectance of the band on that day after radiation correction. 20-30 data are collected every day, and a total of 1040 data are obtained.

[0069]

[0070] Figure 2a-2d It is the average reflectivity of samples collected in different bands during the 12-day data collection period. Figure 2a 550nm is the wavelength of green light. Figure 2b Corresponding to 660nm, Figure 2c Corresponding to 730nm, Figure 2d Corresponding to the 790nm band. From the image trend, it can be seen that the spectral reflectance value range of the 550nm and 660nm bands is close, and the data change trend is slightly similar. The reflectance value range and change trend under the 730nm and 790nm bands are also basically the same. With the change of acquisition time, the reflectance value gradually decreases.

[0071] 1.2.3 Data collection on cotton plant height, number of fruit branches, number of flower buds and determination of topping time

[0072] In order to determine the correlation between cotton canopy spectral information and cotton plant height, fruit branches, flower buds and other indicators, field measurements of cotton plant height, fruit branches and flower buds were conducted: 5 plants with a long distance and fixed positions were randomly selected as sample measurement points in the experimental plot, and the height of these 5 fixed cotton plant samples was measured with a tape measure every day, and the average value was taken as the plant height value of the plot. The number of fruit branches and flower buds of the 5 fixed plants was measured by counting method, and the average value was taken as the number of fruit branches and flower buds of the plot on that day.

[0073] Figure 3a-3cThe schematic diagrams of the measurement of cotton plant height, fruit branches, and flower bud number over time are given respectively. As can be seen from the figure, the cotton plant height, fruit branches, and flower bud number continued to increase from 62cm, 6.5, and 17 on the first day, respectively, until they almost stopped changing after the 7th to 8th day. The test field was toppled on the 8th day. As can be seen from the figure, the various cotton indicators almost stopped changing after topping.

[0074] 1.3 Data processing methods

[0075] 1.3.1 Decision Tree

[0076] The decision tree algorithm is a data mining algorithm with the characteristics of high readability and fast operation speed. The decision tree is constructed by calculating conditional probabilities. The decision tree mainly includes two processes: building the tree structure and pruning the tree. When learning, the training data is used to build a decision tree model based on the principle of minimizing the loss function.

[0077] Common algorithms for building decision trees include: ID3 algorithm and C4.5 algorithm. Decision trees use the maximum likelihood method to calculate probability models. The core of the two algorithms is to apply information gain (ratio) to select features at each node of the decision tree, use a recursive method to build a decision tree, and start from the root node to calculate all possible feature information (ratio) gains for the node, select the feature with the largest information gain (ratio) as the feature of the node, and establish child nodes based on different values ​​of the feature. Then recursively call the above method on the child nodes to build a decision tree until the gains (ratios) of all features are very small or there are no feature positions. Among them, ID3 uses information gain to select features, and C4.5 uses information gain ratio to select features.

[0078] g(D,A)=H(D)-H(D|A) (3)

[0079]

[0080]

[0081] Formula (3) represents information gain, where the training data set is D, the feature is A, and H(D) represents the empirical entropy of data set D. H(D|A) represents the empirical conditional entropy of feature A on the data set. Where n is the number of values ​​of feature A. Formulas (4) and (5) represent information gain ratios. In the input space where the training data set is located, recursively divide each region into two sub-regions and determine the output value on each sub-region to establish a binary decision tree:

[0082] 1.3.2 Decision Tree Algorithm Based on Ensemble Learning

[0083] Ensemble learning methods are widely used in machine learning and statistical learning. Ensemble methods involve independently training different single regression models using the same or different datasets, with different parameters between regression models. Find the average of all single or multiple but classified outputs to obtain the final expected output.

[0084] Random forest is a combination of decision tree and bagging. The bagging algorithm is described as follows:

[0085] Input: training set D = (x1, y1), (x2, y2)....., (x n ,y n ), basic learning algorithm The number of training rounds T;

[0086] Output:

[0087] (1)fort=1,2,..T do

[0088] (2)

[0089] (3)end

[0090] In order to increase randomness, the sampling method with replacement was used to select n samples during the sample collection process.

[0091] (1) Select n samples from the sample set using Bootstrap sampling (sampling with replacement).

[0092] (2) Randomly select k attributes from all attributes and select the best split attribute as a node to build a CART decision tree

[0093] (3) Repeat the above two steps m times to build m CART decision trees.

[0094] (4) These m CARTs form a random forest and decide the data regression value through voting results.

[0095] This application uses Adaboost + decision tree. Adaboost is an adaptive reinforcement learning algorithm, and the algorithm is described as follows:

[0096] Input: training data set T = {(x1,y1),(x2,y2)…,(x n ,y n )},in ,y i ∈y, is the instance space, and y is the set of tags.

[0097] Output: Final classifier: G(x).

[0098] (1) Initialize the weight distribution of training data

[0099]

[0100] (2) For m = 1, 2, ..., M

[0101] (a) Using a weighted distribution D m The training data set is used to learn and obtain the basic regression model

[0102] G m (x):X→y (7)

[0103] (b) Calculate G m (x) Error rate on the training data set

[0104]

[0105] (c) Calculate G m (x i )

[0106]

[0107] (d) Update the weight distribution of the training dataset

[0108] D m01 ={w m+1,1 ,…,w m+1,i ,…,w m+1,N} (10)

[0109]

[0110] Here Z m is the normalization factor

[0111]

[0112] It makes D m+1 becomes a probability distribution.

[0113] (3) Constructing a linear combination of basic regression models

[0114]

[0115] The final regression model is obtained.

[0116]

[0117] 1.3.3 Other multispectral analysis algorithms

[0118] Traditional multispectral analysis methods may include: multiple linear regression, neural network, support vector machine, adaptive competitive weighting, decision tree, etc.

[0119] The multivariate linear method fits the input spectral data and the target data by finding reasonable polynomial parameters. The training time is short, but the fitting effect for linear inseparable data is poor. The neural network uses vector multiplication and nonlinear activation function to multiply the input data with the connection weight and map the result to the nonlinear space. The connection weight is adjusted by the error back propagation method to obtain a better fitting effect. The support vector machine can select the kernel function of the linear or nonlinear kernel to map the data to the high-dimensional space, and extract the high-dimensional features of the data by finding the maximum interval hyperplane. Neural networks and support vector machines integrate feature extraction and regression analysis. The adaptive competitive weighting method is a commonly used analysis method for hyperspectral remote sensing data. The wavelength points with large absolute values ​​of the correlation coefficient of the partial least squares model are selected through the adaptive reweighted sampling technology, and the wavelength points with small weights are removed. The subset with the lowest root mean square error is used to obtain the optimal band subset, so as to achieve the purpose of feature extraction and regression analysis.

[0120] 2 Experiments and results.

[0121] The multispectral data after radiation correction was used as the model input, and a regression prediction model was established for indicators such as cotton plant height, number of fruit branches, and number of flower buds. The regression model selected multiple simple regression methods such as multivariate linear regression, neural network, support vector machine, adaptive competitive weighting method, and decision tree. To ensure the convergence of the model, the algorithm iteration number required was 30,000 times. 920 samples were used as model training samples, and 120 data samples were used as test samples. First, each spectrum was input into the simple regression model to analyze the impact of each spectrum on the prediction results, and then multiple bands were combined for regression analysis based on the regression results. Due to the short time interval of data collection and the small change of cotton plant growth, in order to improve the feature extraction and target fitting ability of the model, the fitting effect of the simple regression model was compared, and the model with good fitting effect and stable performance was selected as the basis for improvement. The bagging and Adaboost ensemble learning algorithms were used to improve the prediction ability of the simple regression model.

[0122] 2.1 Regression model between cotton canopy spectra in different bands and cotton plant height, fruit branches, and number of flower buds

[0123] In this experiment, the spectral reflectance of different bands was first input into a variety of simple regression models, and fitted regression analysis was performed with cotton plant height, number of fruit branches, and number of flower buds. The fitting results are shown in Table 1.

[0124] Table 1 Regression table of cotton canopy and cotton plant height, flower buds and fruit branches in different bands

[0125]

[0126]

[0127]

[0128] From the fitting results, the 730nm band has the best fitting effect on plant height, number of flower buds and number of fruit branches. The 790nm band is second, and the fitting coefficients with the three targets can reach above 0.65. The 550nm band has a poor fitting effect, and the correlation coefficients for each indicator are all below 0.3. The 660nm band has the worst fitting effect, and the fitting correlation coefficients for each indicator are all below 0.1. From the regression results of a single band, it can be seen that the regression effects of the 790 and 730nm bands on cotton plant height, number of flower buds and number of fruit branches are better than those of the 550 and 660nm bands. The regression effect of 550nm is slightly better than that of 660nm.

[0129] From the perspective of regression methods, in the regression model of a single band, both neural networks and decision trees have good regression effects. The regression coefficient and prediction mean square error of the neural network are slightly higher than those of the decision tree, indicating that the neural network has a high degree of fit to the data but the effect is unstable. The support vector machine method is second to the above two methods in terms of regression coefficient, and the multivariate linear regression effect is the worst, which shows that the characteristics of multispectral data are nonlinear and need to be extracted by nonlinear methods.

[0130] 2.2 Regression model of traditional vegetation index and cotton plant height, fruit branches and flower bud number

[0131] A variety of traditional vegetation indices were used to regress the plant height, flower buds and fruit branches of cotton. Due to the experimental equipment used in this paper, vegetation indices related to 660, 730, and 790 nm spectra were selected for regression. Finally, the difference index (DI), difference vegetation index (DVI), red edge chlorophyll index (CI red edge ), Normalized Difference Vegetation Index (NDVI), Green Normalized Difference Vegetation Index (GNDVI), and Triangular Vegetation Index (TVI). The calculation formulas of several indices are as follows: In the formula, R represents the spectrum, and the subscript number represents the band of specified length. Since the spectral resolution error of the spectral instrument is ±30nm, R 800 The 790nm band can be used instead, R 680 , R 670 660nm can be used instead, R 720 , R 750 Use 730nm instead. The regression effect is shown in Table 2.

[0132] DI=R 800 -R550 (15)

[0133] DVI=R 800 -R 680 (16)

[0134]

[0135] NDVI=(R 780 -R 670 ) / (R 780 +R 670 ) (18)

[0136] GNDVI=(R 800 -R 550 ) / (R 800 +R 550 ) (19)

[0137] TVI=0.5*(120*(R 750 -R 550 )-200*(R 670 -R 550 )) (20)

[0138] Table 2 Regression table of traditional vegetation index and cotton height, flower buds and fruit branches

[0139]

[0140]

[0141]

[0142] From the results in Table 2, we can see that DVI and CI rededge The regression effects of the two traditional vegetation indices on cotton plant height, flower buds and fruit branches were relatively stable and the regression fit was high. The fitting degree of other indices on the three indices was unstable, and the changes in fitting indices and fitting methods had a greater impact on the fitting effect. The spectral combination data had a better regression effect than the single spectrum fitting, among which the vegetation index designed in the 730 and 790 nm bands had a better fitting effect, which was the same as the regression comparison effect of a single spectrum.

[0143] 2.3 Multispectral regression model of cotton canopy and cotton plant height, fruit branches and bud number under different band combinations

[0144] Table 3 Regression table of different band combinations and cotton height, flower buds and fruit branches

[0145]

[0146]

[0147]

[0148] This paper uses the band combination method to combine multiple bands, input them into the regressor, and analyze the results. The regression effect is shown in Table 3. In Table 3, the bands with better regression effects are superimposed to form a new band combination and input into the regressor for analysis. From the analysis results, it can be seen that the 730+790nm superposition has a better regression effect than a single band, and the regression coefficient of cotton plant height, flower buds, and fruit branches can reach more than 0.80. The combination of 550+730+790nm is slightly better than the 730+790nm band combination, indicating that 550nm has a certain positive effect on the fitting results. The full-band spectrum (550+660+730+790)nm is used to regress the relevant indicators. From the regression effect, the full-band combination is degraded compared with the 730+790nm combination and the 550+730+790nm band combination. It shows that the addition of the 660nm band has a negative effect on the regression.

[0149] From the perspective of regression methods, both neural networks and decision trees perform well. Decision trees have higher correlation coefficients and smaller mean square errors than neural networks, and the regression effects on multiple indicators are relatively stable. The performance of support vector machines and multivariate linear regression models is relatively average.

[0150] 2.3 Combined band cotton topping index regression model based on improved decision tree

[0151] Table 4 Regression of cotton canopy multispectral and cotton plant height

[0152]

[0153]

[0154] From the perspective of band combination, the 730+790nm band combination, the 550+730+790nm band combination and the full band data were fitted and analyzed using the random forest and Adaboost+decision tree methods, respectively. The fitting results are shown in Table 3. The 550+730+790nm band combination has a better regression effect than the 730+790nm band combination. Since the 550nm band itself has a positive correlation with the cotton topping period indicators, it can supplement the 730+790nm information. The regression effect of the full band is weaker than that of the 550+730+790nm band combination. This is because the 660nm band has a poor regression effect on cotton-related indicators, which interferes with the regression effect.

[0155] From the methodological point of view, in the simple regression model, the decision tree model and the neural network have good performance in the regression correlation coefficient, and the decision tree has a smaller prediction mean square error, which means that the prediction effect of the decision tree is more stable. Therefore, this paper improves the method based on the decision tree, and uses boosting and Adaboost methods to improve the decision tree model and the regression model. The boosting method improves the model's fitting ability by establishing different subsets and assigning different weights to different subsets. The Adaboost method improves the model's feature extraction ability by establishing multiple regression models and assigning different weights to the regression models. From the comparative analysis of the regression effects, it can be seen that the regression effect of Adaboost+decision tree is better, the correlation coefficient is higher, the prediction mean square error is smaller, and it has stronger feature extraction ability.

[0156] The 550+730+790nm band combination has the highest fitting degree for cotton plant height, number of flower buds, and number of fruit branches. Among them, the Adaboost+decision tree method has the largest fitting correlation coefficient, which is 0.96, 0.84, and 0.97, respectively, and the mean square error is the smallest, which is 2.4 cm, 0.49, and 0.54, respectively. For the 730+790nm band combination, the regression coefficients of the decision tree and Adaboost+decision tree methods are the same, and the correlation coefficients for plant height, number of flower buds, and number of fruit branches are 0.90, 0.91, and 0.95, respectively. However, the prediction mean square error of Adaboost+decision tree is smaller, which is 3.86 cm, 0.95, and 0.99, respectively. The regression effect of the whole band is shown in Table 3, among which the fitting coefficients of the Adaboost+decision tree method are 0.93, 0.90 and 0.95 respectively, and the prediction mean square error is the smallest, which is 2.03cm, 0.97 and 0.52 respectively. The prediction effect of the combination of Adaboost+decision tree and red light band data on cotton plant height, number of fruit branches and number of flower buds is shown in Table 3. Figure 4a-4c In summary, using the 550+730+790nm band combined data and the Adaboost+decision tree method to perform regression analysis on the data can obtain the best prediction effect of cotton plant height, flower buds and fruit branches.

[0157] 2.4 Cotton topping time determination test

[0158] According to the "Guiding Opinions on the Technical Guidance of Cotton Mechanization Production in the Northwest Inland Cotton Region" issued by the Ministry of Agriculture and Rural Affairs and the requirements for cotton topping time in related cotton planting agronomy: determine the appropriate topping time based on factors such as cotton growth, plant height and number of fruit branches, topping early in time, and determine plant monitoring indicators based on early maturity. Based on the indicators for determining the topping period in the guiding opinions, this paper quantified the topping indicators of Xinlu Zao No. 68 cotton in Shihezi, Xinjiang, through field research and experimental data analysis, and determined that the topping time determination indicators for this variety are 65-70cm plant height, 9-11 fruit branches, and 18-21 buds when cotton is topping.

[0159] On this basis, the plant height, fruit branches, and flower bud number indicators predicted by the 550+730+790nm band combination through neural network, decision tree, random forest, and Adaboost+decision tree methods were used to judge whether to top the cotton. The predicted topping results were compared with the measured topping results, and the statistical judgment accuracy is shown in Table 4. As can be seen from Table 4, the 550+730+790nm band combination and the Adaboost+decision tree method have better effects on the determination of cotton topping time than other band combinations and other methods.

[0160] Table 5 Prediction accuracy of cotton topping based on regression model

[0161]

[0162]

[0163] According to the experimental results, this paper selected the combination of green light + red light + near infrared (550 + 730 + 790nm) combined with the decision tree method to obtain the best regression model. The prediction correlation coefficients for plant height, fruit branches, and flower bud number were 0.91, 0.88, and 0.96, respectively, and the prediction mean errors were 3.91cm, 3.98, and 1.75, respectively. The model prediction results were input into the topping judgment model, and the judgment accuracy was 85.23%. The ensemble learning method can improve the feature extraction ability of the model and improve the fitting ability of the model. Boosting and Adaboost methods are common ensemble learning methods. Boosting and Adaboost methods are used to improve the decision tree. From the improved results, Adaboost + decision tree has a better fitting effect. The prediction correlation coefficients for plant height, fruit branches, and flower bud number are 0.96, 0.84, and 0.97, and the prediction mean square errors are 2.4cm, 0.49, and 0.54, respectively. The data predicted by the model were input into the topping determination model, and the determination accuracy was 94.03%.

[0164] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A method for predicting the growth status of cotton plants during the topping period based on multispectral remote sensing by unmanned aerial vehicles, characterized in that It includes the following steps: S1. Use a multispectral camera to collect multispectral information of the cotton canopy before and after the topping operation; collect crop indicators related to the topping time strategy, and the crop indicators related to the topping time strategy include: cotton plant height, number of fruit branches, and number of flower buds; S2. Construct a multispectral regression analysis model of cotton canopy multispectral information-crop index; the multispectral regression analysis model is constructed by combining Adaboost + decision tree. The algorithm steps of Adaboost + decision tree are as follows: Input: training data set T = {(x1, y1), (x2, y2) ..., (x n ,y n )},in y i ∈y, X is the instance space, y is the label set; Output: Final classifier: G(x); (1) Initialize the weight distribution of training data: Among them, w 1i is the weight of the initialization data, N is the number of training data; (2) For m = 1, 2, ..., M (a) Using a weighted distribution D m The training data set is used to learn the basic regression model: G m (x):X→y G m is the final classifier; (b) Calculate G m (x) Error rate e on the training data set m : w mi represents the weight of the i-th instance in the m-th round, Description m (x) The regression error rate on the weighted training dataset is G m The sum of the difference weights between the predicted value and the actual value of (x), from which we can see the weight distribution D of the data m With the basic classifier G m (x) the relationship between the classification error rate; (c) Calculate G m Coefficient distribution of (x)α m : (d) Update the weight distribution of the training data set: D m+1 ={w m+1,1 ,...,w m+1,i ,...,w m+1,N } Here Z m is the normalization factor: It makes D m+1 becomes a probability distribution; (3) Construct a linear combination of basic regression models: The final regression model is obtained: S3. Using the multispectral information of the specific band combination collected from the cotton field to be predicted as input data, the crop index is obtained through the multispectral regression analysis model; the specific band combination is 550nm, 730nm, and 790nm.

2. The method according to claim 1, characterized in that In S1, a drone equipped with a multispectral camera is used to obtain multispectral information of the cotton canopy.

3. The method according to claim 1, characterized in that In S1, a standard plate of fixed size is placed in the cotton field to be predicted, and the spectrum of the standard correction area is obtained as the radiation correction data of the remote sensing data. The correction formula is: Where I represents the average spectral value of the region of interest in a certain band, W represents the spectral mean of the standard white plate correction area of ​​the band on that day, B represents the pixel mean of the band when the lens is covered on that day, and CI is the spectral reflectance of the band on that day after radiation correction.

Images