Method for determining the time of wheat fusarium rust prevention and plant protection operation based on UAV multispectral remote sensing

Abstract

The present invention discloses a method for determining the time of wheat fusarium head blight prevention and plant protection operation based on multispectral remote sensing of unmanned aerial vehicles, which includes: S1, using a multispectral camera to collect canopy spectral information of wheat heading and flowering period; S2, based on the canopy spectral information, using a one-dimensional convolutional neural network to extract characteristic data affecting heading rate; S3, based on the characteristic data, using a decision tree model to construct a regression analysis model of wheat heading rate-canopy spectral information; S4, using the multispectral information of a specific band combination collected from a wheat field to be predicted as input data, obtaining the wheat ear situation through the regression analysis model obtained by S3, and according to the requirements for the fusarium head blight prevention and control period in the "Guidance on Technical Prevention and Control of Wheat Fusarium Head Blight in 2016", when the heading rate is greater than 0.9, it is determined that the wheat is already in the fusarium head blight prevention and control period. The scheme provided by the present application can provide crop data support for the determination of unified prevention and control time and strategy for wheat heading period.

Description

Technical Field

[0001] The present invention belongs to the field of smart agriculture, and in particular is a method for determining the time of wheat fusarium wilt prevention and plant protection operations based on unmanned aerial vehicle multispectral remote sensing. Background Art

[0002] Wheat is an important food crop, and ensuring wheat yield and quality is an important part of ensuring food security. Fusarium head blight is a worldwide wheat disease that seriously affects wheat yield. According to the "2016 Wheat Fusarium Head Blight Prevention and Control Technical Guidance" issued by the Ministry of Agriculture and Rural Affairs and related agronomic requirements, the overall prevention and control idea of ​​"prevention first" should be adhered to. Due to the characteristics of Fusarium head blight being highly contagious, difficult to cure, and having a great impact on yield, Fusarium head blight is focused on prevention and control, while taking into account the prevention and control of other pests and diseases such as powdery mildew, aphids, and armyworms. The opinions point out that the best period for preventing and controlling the occurrence of wheat Fusarium head blight is from the beginning of wheat heading to the early flowering stage. The time of pesticide spraying has a great influence on the prevention effect of the disease. The prevention and control operation time of Fusarium head blight is closely related to the wheat heading rate. In actual production operations, the monitoring of wheat heading rate mainly relies on manual inspections, which are inefficient and highly subjective. Therefore, it is of great significance to quickly and accurately grasp the wheat heading rate for the prevention and control effect of wheat Fusarium head blight. In addition, research and experiments have shown that mastering the crop operation time window is of great significance to the planning of plant protection machinery operation strategies and improving the effectiveness and efficiency of plant protection operations.

[0003] At present, the research on monitoring physiological indicators of wheat is mainly focused on the jointing stage, flowering stage or filling stage of wheat, and the monitoring targets are indicators such as chlorophyll and leaf area index of wheat. Some studies have also focused on monitoring after the onset of wheat fusarium rust, but there is a lack of research on monitoring during the prevention of wheat fusarium rust during the heading and flowering stage. Low-altitude remote sensing technology based on unmanned aircraft has the advantages of high maneuverability, simple and safe operation, and is applied to crop growth monitoring. Multispectral cameras can obtain more band information than ordinary digital cameras, and higher spectral cameras have a more economical price. This hardware has more advantages in monitoring effect and economic cost. Common multispectral information processing methods include: neural network, support vector machine, decision tree and other simple feature extraction methods, all of which can achieve good target fitting accuracy. In the past, the experimental process with chlorophyll, leaf area and nitrogen changes as the research targets had a long time interval for data collection, and the crop growth state had a large change, so the feature extraction ability of the fitting model was required to be low. In the heading and flowering stage of wheat, except for the rapid change of heading rate, other features did not change significantly, so there were high requirements for the feature extraction and expression ability of the regression model. In summary, there are problems in the prevention of wheat fusarium wilt: the information collection efficiency for determining the time of plant protection operations is low and highly subjective. Summary of the invention

[0004] In view of the problems existing in the background technology, this application intends to use a drone equipped with a multispectral camera to collect canopy spectral information during the heading and flowering period of wheat, and on this basis, design a one-dimensional convolutional neural network to extract features from the spectral information and fit the wheat heading rate. According to the requirements of wheat fusarium head blight prevention and plant protection time, a wheat fusarium head blight prevention and plant protection operation discrimination model is established to provide a method for quickly distinguishing crop status for determining the unified prevention and control time of wheat heading period and the operation strategy of intelligent plant protection equipment.

[0005] Technical solution:

[0006] A method for determining the time of wheat fusarium rust prevention and plant protection operation based on unmanned aerial vehicle multispectral remote sensing, comprising:

[0007] S1. Use a multispectral camera to collect canopy spectral information during the heading and flowering period of wheat;

[0008] S2, based on canopy spectral information, a one-dimensional convolutional neural network was used to extract characteristic data affecting heading rate;

[0009] S3. Based on the characteristic data, a regression analysis model of wheat heading rate-canopy spectral information was constructed using a decision tree model;

[0010] S4. Using the multispectral information of the specific band combination collected from the wheat field to be predicted as input data, the wheat ear emergence situation is obtained through the regression analysis model obtained in S3. According to the requirements for the period of fusarium head blight prevention and control in the "Technical Guidelines for the Prevention and Control of Wheat Fusarium Head Blight in 2016", when the heading rate is greater than 0.9, it is determined that the wheat is in the period of fusarium head blight prevention and control.

[0011] Preferably, in S1, a drone equipped with a multispectral camera is used to collect canopy spectral information during the heading and flowering period of wheat.

[0012] Preferably, in S1, a standard plate of a fixed size is placed in the wheat 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:

[0013]

[0014] 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.

[0015] Preferably, in S2, the feature data is full-volume base layer data calculated by a one-dimensional convolutional neural network.

[0016] Preferably, in S2, the step of acquiring the feature data is:

[0017] S2-1. Construct a one-dimensional convolutional neural network, including 5 convolutional layers. Each convolutional layer uses the LeakyRelu function as the activation function. The last three layers add a dropout module to improve the training speed and generalization performance of the network. The dropout probability is set to 0.1. The features extracted by the third layer of the network are combined with the features extracted by the fifth layer of the network and then input into the fully connected layer. The convolution kernels in the convolutional neural network are all 1-dimensional, and the convolution kernel weights are initially randomized. Since the input dimension is 4, the size of the convolution kernel is 1, the step size is 1, and the 'same' method is used to perform convolution operations on the data.

[0018] S2-2. After the one-dimensional convolutional neural network training is completed, the output in the fully connected layer is extracted as feature data.

[0019] Preferably, in S3, the feature data is input into a decision tree, and the decision tree is used to perform regression analysis on the feature information to fit the wheat heading rate.

[0020] Specifically, in S2-1, the training objective function of the one-dimensional convolutional neural network is:

[0021]

[0022] In the formula, n is the number of data in the entire data set, p i is the predicted value, a i is the true value.

[0023] Specifically, in S2-1, the Adam method is used to optimize the weights of the convolution kernel, where the learning rate is lr=0.002, the decay rate is 1e-9, the momentum is 0.5, and the epoch is 1000.

[0024] Specifically, in S4, the specific wavelength band combination is 550nm, 660nm, 730nm and 790nm.

[0025] Beneficial Effects of the Invention

[0026] The solution provided in this application can provide crop data support for determining the time and strategy for unified prevention and control during the wheat heading period.

[0027] This application designs a one-dimensional convolutional neural network + decision tree structure to process wheat canopy spectral information. The fitting correlation coefficient of the canopy multi-spectrum processed by the fitting model to the heading rate can reach 0.95, and the prediction mean error (RMSEP) is 0.24. The prediction result has an accuracy rate of 97.50% for the unified prevention and control time of wheat fusarium.

[0028] The 550+660+730+790nm band combination has a better fitting effect on the prediction of wheat heading rate than the traditional vegetation index. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Schematic diagram of spectral reflectance of wheat in the 550nm band

[0030] Figure 2 Schematic diagram of spectral reflectance of wheat in the 660nm band

[0031] Figure 3 Schematic diagram of spectral reflectance of wheat in the 730nm band

[0032] Figure 4 Schematic diagram of spectral reflectance of wheat in the 790nm band

[0033] Figure 5 This is a schematic diagram of the heading rate change of Zhenmai 7 wheat

[0034] Figure 6 This is a schematic diagram of the heading rate change of Yangmai No. 12 wheat

[0035] Figure 7 This is a schematic diagram of the heading rate change of Yangmai 16 wheat

[0036] Figure 8 Schematic diagram of the convolutional neural network structure

[0037] Fig. 9 Schematic diagram of the convergence process of neural network training

[0038] Fig.10 This is a test graph of wheat heading rate fitting based on the convolutional neural network fitting method. DETAILED DESCRIPTION

[0039] The present invention will be further described below in conjunction with embodiments, but the protection scope of the present invention is not limited thereto:

[0040] 1 Materials and methods

[0041] 1.1 Overview of the experimental area and experimental materials

[0042] The experimental data collection site is located in Taiping Village, Qixia District, Nanjing City, Jiangsu Province. Taiping Village is located in the northeast of Nanjing and the lower reaches of the Yangtze River. The terrain of the experimental area is a plain, with a northern subtropical humid climate, four distinct seasons, rain and heat in the same season, sufficient sunlight, and an average annual precipitation of 1090.4 mm. The average annual temperature is 15.4℃, and the frost-free period is 237 days. The soil type is yellow-brown soil. The wheat varieties collected in this article are Zhenmai No. 7, Yangmai 12, and Yangmai 16, with one plot for each variety, and the size of each plot is about 0.3 hectares. The collection time was from April 11 to April 20, 2021, covering wheat with a heading rate of 0-100%, and a total of 720 data were obtained for 10 consecutive days.

[0043] 1.2 Data collection and preprocessing

[0044] 1.2.1 Multispectral Data Collection

[0045] The embodiment uses the XAG drone to collect remote sensing data of wheat canopy. The target plot is selected in the smartphone system. The drone flies according to the route automatically planned by the system. The drone flies at an altitude of 9m and a speed of 3m / s. The multispectral camera has 20 million pixels, an image resolution of 3863*3648 pixels, and a camera weight of 0.85kg. The multispectral camera covers four bands, and the wavelengths of the corresponding central bands are 550, 660, 735, and 790nm. A 50*50cm standard whiteboard is placed on each operating plot as radiation correction data for later remote sensing data.

[0046] 1.2.2 Multispectral Data Preprocessing

[0047] (The solution claimed in this application) After the drone hyperspectral remote sensing data is acquired, it needs to be processed. The data processing mainly includes two parts: (1) Extraction of the region of interest, using a 50*50 pixel rectangular box to manually extract on the multispectral image, and average the values ​​in the region of interest as a sample data. (2) Radiation correction, using a 30*30 pixel rectangular box to calibrate the position of the standard white board, and calculate its average value as the white standard correction value. On this basis, radiation correction is performed using formula (1), where W is the standard white board spectrum mean value in this band on that day, I is the sample spectrum mean value in this band on that day, and B is the pixel mean value in this band when the lens is covered.

[0048]

[0049] The average multispectral reflectance of different varieties of wheat on the same day is as follows Figure 1-Figure 4 As shown in the figure, the reflectance values ​​of 730 and 790nm bands are larger than those of 550 and 660nm, and the spectral reflectance change trends of the four bands are different. From the perspective of varieties, Zhenmai No. 7 and Yangmai No. 16 have similar change trends in each band, while Yangmai No. 12 has a slightly different change trend from the previous two.

[0050] (Comparative example) In addition to directly inputting the information of each band into the fitting model, it is also necessary to study the transformation of the values ​​of two or more bands to form a vegetation index, so as to better highlight the changes in crop characteristics and better fit the target. This paper selects the difference index (DI), difference vegetation index (DVI), red edge chlorophyll index (CIrededge), normalized difference vegetation index (NDVI) and green normalized difference vegetation index (GNDVI), triangle vegetation index (TVI) according to the collected spectral wavelength.

[27] , these five vegetation index information are fitted to the wheat heading rate. The calculation formulas of several indices are as follows (2)-(7): 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.

[0051] DI=R 800 -R 550 (2)

[0052] DVI=R 800 -R 680 (3)

[0053] NDVI=(R 780 -R 670 ) / (R 780 +R 670 ) (4)

[0054]

[0055] GNDVI=(R 800 -R 550 ) / (R 800 +R 550 ) (6)

[0056] TVI=0.5*(120*(R 750 -R 550 )-200*(R 670 -R 550 )) (7)

[0057] 1.2.3 Wheat Heading Rate Data Acquisition

[0058] Wheat heading rate: In this experiment, data of three varieties of wheat were collected, one plot for each variety, three sampling points were randomly selected in each plot, and 20 wheat plants were selected at each sampling point as the area of ​​interest. The number of wheat ears in the area of ​​interest was counted using the counting method, and the average heading rate of the three sampling points was taken as the heading rate of the plot. The changes in the heading rates of the three varieties are expressed as follows: Figure 5-Figure 7As shown in the figure, it can be seen that the heading rate range of Zhenmai No. 7 is [0.08, 0.86], the heading rate of Yangmai No. 12 is [0.27, 1], and the heading rate range of Yangmai No. 16 is [0.3, 1]. The heading rate change trends of each variety are slightly different. The heading rate of Zhenmai No. 7 increases rapidly in the first 8 days, and the heading rate increases slowly from the 8th to the 10th day. The heading rate of Yangmai No. 12 has been in a rapid growth state in the first 3 days, and the heading rate increases rapidly from 0.3 to 0.98, and increases from 0.98 to 1 from the 3rd to the 5th day, and then remains at 1. The heading rate of Yangmai No. 16 from the 1st day to the 3rd day ranges from [0.3, 0.5], and from the 3rd day to the eighth day ranges from [0.5, 1]. The change trend is relatively gentle, and the heading rate on the 9th and 10th days no longer changes.

[0059] From the trend of heading rate, Zhenmai 7 and Yangmai 16 have similar trends, and it took 8-9 days from partial heading to full heading, while Zhenmai 12 has a fast heading speed, and it only took 5 days from partial heading to full heading. Comparing the spectral change patterns of the three varieties of wheat, Zhenmai 7 and Yangmai 16 have similar spectral change trends to Yangmai 12, and the heading trend is also similar to Yangmai 12. Therefore, it can be inferred that there is a certain correlation between the multi-spectral information of the wheat canopy and the heading rate.

[0060] 1.3 Fitting method

[0061] A total of 720 multispectral data and corresponding heading rate labels were obtained in the experiment. 120 of them were randomly selected as test data, and the remaining 600 were used as training data. The size of each input data is 4*1 dimension, and the output data is 1*1.

[0062] 1.3.1 One-dimensional convolutional neural network + decision tree model (the solution claimed in this application)

[0063] Since the crop state changes little during the wheat heading period, higher requirements are placed on the feature extraction and expression capabilities of the convolutional neural network. The first few layers of convolution in the convolutional network can extract surface features of information such as color and texture, and the deep layer of the network can extract abstract features of information. By stacking multiple layers, better feature extraction capabilities can be obtained. Based on this principle, combined with Figure 8When designing the network structure, this paper combines the features extracted by the third layer of the network with the features extracted by the fifth layer of the network, and then inputs them into the fully connected layer. The convolution kernels in the convolutional neural network are all 1-dimensional, and the convolution kernel weights are initially calculated using the random number method. Since the input dimension is 4, the size of the convolution kernel is 1, the step size is 1, and the 'same' method is used to perform convolution operations on the data. The neural network has a total of 5 convolutional layers, and each convolutional layer uses the LeakyRelu function as the activation function. The last three layers add a dropout module to improve the training speed and generalization performance of the network. The probability of dropout is set to 0.1. The training objective function of the convolutional neural network uses mean_squared_logarithmic_error, and its formula is shown in (8). Where n is the number of data in the entire data set, p i is the predicted value, a i The Adam method is used to optimize the weights of the convolution kernel, where the learning rate is lr = 0.002, the decay rate is 1e-9, the momentum is 0.5, and the epoch is 1000.

[0064]

[0065] Fig. 9 The convergence process of neural network training is given. After the network is trained, the output of the fully connected layer is extracted as features, the features are input into the decision tree, and the decision tree is used to perform regression analysis on the feature information to fit the wheat heading rate.

[0066] 1.3.2 Neural Network, Support Vector Machine and Decision Tree Method (Comparative Example)

[0067] Neural networks use nonlinear transformation methods to find appropriate parameters in the input space and solution space to achieve the purpose of information transformation. When the network is initialized, the connection weights between neurons at different levels are determined using a random number method, and the distance between the network calculation value and the target value is calculated using the minimum mean square error as the objective function. The error is propagated to each neuron connection weight using the reverse error propagation method, and the connection weight is adjusted according to the gradient descent direction and step size.

[0068] The experiment in this paper found that the 4-layer network structure, namely: the number of neurons in the input layer is 4, the number of neurons in the first hidden layer is 30, the number of neurons in the second hidden layer is 15, and the number of neurons in the output layer is 1, is the best experimental structure.

[0069] Support vector machine is a generalized linear classifier that performs binary classification of data in a supervised learning manner. Its boundary decision is to solve the maximum margin hyperplane for learning samples. Since most data are nonlinearly separable, the kernel function method is used to map the data into a high-dimensional space. This paper uses the Gaussian kernel to map the data into a high-dimensional space. The objective function uses the minimum mean square error and determines the parameters of the support vector machine by finding the maximum error segmentation plane method.

[0070] The conditional probability distribution of a decision tree under given feature conditions. A set of classification or regression rules is summarized through the training data set. A decision tree is a tree structure. Each node in the tree represents an object, and each fork represents a possible attribute value. Each leaf node corresponds to the value of the object represented by the path from the root node to the leaf node. The nodes mainly include: decision nodes, chance nodes, and endpoints. A tree structure with good fitting degree and generalization performance is obtained through pruning and learning.

[0071] Both neural networks and decision trees require repeated iterations to determine the connection weights or tree structures. The number of iterations in this article is 10,000.

[0072] 2 Experimental results

[0073] 2.1 Wheat Heading Rate Estimation Model Based on One-Dimensional Convolutional Neural Network

[0074] First, all the band information is used, that is, the dimension of each data is 1*4, and it is directly input into the convolutional neural network, neural network, support vector machine and decision tree models.

[0075] Table 1 Fitting effect of full-band information input

[0076]

[0077]

[0078] From the fitting results, it can be seen that the convolutional neural network has the best fitting effect, with a determination coefficient of 0.95 and a prediction mean square error of 0.24. It can obtain a better fitting effect than using the decision tree alone, and its determination coefficient is increased by 0.12. The fitting determination coefficients of the neural network, support vector machine and decision tree are all around 0.8, among which the determination coefficient of the decision tree is higher, indicating that the decision tree has better feature extraction and fitting expression capabilities for the data than the other two methods. In summary, the convolutional neural network has a stronger feature extraction capability than other methods. After extracting features, combining it with the decision tree for regression analysis can obtain a better fitting effect.

[0079] Fig.10This is a schematic diagram of the predicted and measured values ​​of wheat multispectral heading rate prediction using the convolutional neural network + decision tree method. It can be seen from the figure that, overall, the predicted and measured values ​​have a good fitting effect, and the difference between the predicted and actual values ​​is within 0.1, but the fitting degree difference for a small number of data is 0.5 or above. In practical applications, the prediction accuracy of the results caused by individual prediction errors can be reduced by selecting multiple regions of interest on a remote sensing image.

[0080] 2.2 Wheat canopy spectrum and heading rate estimation model based on vegetation index (comparative example)

[0081] The difference index (DI), difference vegetation index (DVI), red edge chlorophyll index (CIrededge), normalized difference vegetation index (NDVI), green normalized difference vegetation index (GNDVI), triangle vegetation index (TVI) and wheat heading rate were regressed, and the regression model used neural network, support vector machine and decision tree. The regression results are shown in Table 2.

[0082] Table 2 Fitting effect of traditional vegetation index and wheat heading rate

[0083]

[0084]

[0085] Regression effect analysis: From the perspective of vegetation index, TVI has the best performance, with a determination coefficient of 0.85 and a CI of rededge The performance is second. The TVI determination coefficient is 0.76. TVI is calculated by weighting the information of three bands, involving red edge, red light and green light. rededge The index is mainly calculated from the red edge and infrared bands. From a methodological point of view, decision trees generally have the best performance results compared to support vector machines and neural networks. rededge In the results, the performance of the support vector machine is better than that of the decision tree. The coefficient of determination of the decision tree is 0.04 lower than that of the support vector machine. However, its performance is still not as good as that of the convolutional neural network + decision tree method, which proves that the convolutional neural network has good feature extraction capabilities.

[0086] 2.3 Model for determining the time of wheat fusarium rust prevention and plant protection operations

[0087] This article is based on the provisions of the "2016 Wheat Fusarium Disease Prevention and Control Technical Guidance" issued by the Ministry of Agriculture and Rural Affairs of the People's Republic of China on the spraying time for Fusarium disease prevention and control: the best period to prevent and control the damage caused by wheat fusarium disease is from the beginning of wheat heading to the early flowering period, and "spraying pesticides when flowers bloom" can achieve twice the result with half the effort. Combined with the actual planting experience of farmers, it is determined that the fusarium disease prevention and control operation period is when the heading rate of the entire field reaches 0.9 or above.

[0088] In order to determine the significance and effectiveness of monitoring the heading rate of wheat in this paper, this paper uses a combination of multiple bands to input the combined band data into the convolutional neural network + decision tree monitoring model, and judges the heading rate monitored by the model. If the heading rate is greater than 0.9, it is judged that the wheat is already in the prevention and control period, otherwise it is not in the prevention and control period. The judgment result is compared with the judgment result of the actual heading rate to calculate the accuracy of the prevention and control period. In this paper, each method and band combination information are input into the model for result prediction, and the prediction result is used to predict the prevention and control time, and the judgment accuracy is counted.

[0089] Table 3 Comparison of the accuracy of fusarium head blight control time based on different monitoring models

[0090]

[0091]

[0092] As can be seen from Table 3, from the method point of view, the convolutional neural network + decision tree model has a better monitoring effect on the heading rate of wheat than other methods, indicating that the convolutional neural network has a better feature extraction ability. Among other methods, the decision tree and the support vector machine have good judgment accuracy. Overall, the prediction accuracy of the decision tree has a higher accuracy. From the input band point of view, the fitting effect of the full-band input has a better prediction accuracy than the traditional vegetation index method. Among the traditional vegetation indices, the TVI index has a better prediction accuracy than other indices. The reason is that other traditional vegetation indices only contain two band information such as formulas (2) to (6), while TVI contains three bands such as formula (7). However, in the fitting model established by the same method, its prediction accuracy is lower than that of the full-band information input, indicating that all band information has a better prediction accuracy for topping time judgment.

[0093] In summary, UAVs were used to collect multispectral information of wheat, combined band information was selected, and a fitting model was established using the convolutional neural network + decision tree method to fit the heading rate. The heading rate output by the fitting model was used to determine the topping time, and the determination accuracy could reach 97.50%, obtaining the best fitting effect.

[0094] 4 Conclusion

[0095] (1) The 550+660+730+790nm band combination has a better fitting effect on the prediction of wheat heading rate than the traditional vegetation index.

[0096] (2) A one-dimensional convolutional neural network + decision tree structure was designed to process wheat canopy spectral information. The fitting correlation coefficient of the canopy multi-spectrum processed by the fitting model to the heading rate can reach 0.95, and the prediction mean error (RMSEP) is 0.24. The prediction results have an accuracy rate of 97.50% for the unified prevention and control time of wheat fusarium head blight.

[0097] The data collection and data processing model in this paper can provide crop data support for determining the time and strategy of unified prevention and control during the heading period of wheat.

[0098] 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 determining the time of wheat fusarium rust prevention and plant protection operation based on unmanned aerial vehicle multispectral remote sensing, characterized in that It includes: S1. Use a drone equipped with a multispectral camera to collect canopy spectral information during the heading and flowering period of wheat; S2, based on canopy spectral information, a one-dimensional convolutional neural network was used to extract characteristic data affecting heading rate; S3. Based on the characteristic data, a regression analysis model of wheat heading rate-canopy spectral information was constructed using a decision tree model; S4. Using the multispectral information of the specific band combination collected from the wheat field to be predicted as input data, the wheat heading situation is obtained through the regression analysis model obtained in S3. When the heading rate is greater than 0.9, it is determined that the wheat is in the prevention and control period of ergot disease.

2. The method according to claim 1, characterized in that In S1, a standard plate of fixed size is placed in the wheat 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.

3. The method according to claim 1, characterized in that In S2, the steps for obtaining feature data are: S2-1. Construct a one-dimensional convolutional neural network, including 5 convolutional layers. Each convolutional layer uses the LeakyRelu function as the activation function. The last three layers add a dropout module to improve the training speed and generalization performance of the network. The dropout probability is set to 0.

1. The features extracted by the third layer of the network are combined with the features extracted by the fifth layer of the network and then input into the fully connected layer. The convolution kernels in the convolutional neural network are all 1-dimensional, and the convolution kernel weights are initially randomized. Since the input dimension is 4, the size of the convolution kernel is 1, the step size is 1, and the 'same' method is used to perform convolution operations on the data. S2-2. After the one-dimensional convolutional neural network training is completed, the output in the fully connected layer is extracted as feature data.

4. The method according to claim 1, characterized in that In S3, the feature data is input into the decision tree, and the decision tree is used to perform regression analysis on the feature information to fit the wheat heading rate.

5. The method according to claim 3, characterized in that In S2-1, the training objective function of the one-dimensional convolutional neural network is: In the formula, n is the number of data in the entire data set, p i is the predicted value, a i is the true value.

6. The method according to claim 3, characterized in that In S2-1, the Adam method is used to optimize the weights of the convolution kernel, where the learning rate is lr=0.002, the decay rate is 1e-9, the momentum is 0.5, and the epoch is 1000.

7. The method according to claim 1, characterized in that In S4, the specific wavelength band combination is 550nm, 660nm, 730nm and 790nm.

Images