Pest identification or monitoring method based on hyperspectrum

Abstract

The invention relates to the field of crop pest monitoring, in particular to a method for identifying or monitoring tetranychid damage based on hyperspectrum. In order to solve the problems that the tetranychus urticae is small in size, the harmful symptoms are not easy to discover, and detection and monitoring are difficult, the method has the advantages of a traditional imaging technology and a spectrum technology based on the characteristic that a hyperspectral imaging technology has an image-spectrum unification function, hyperspectral information of plant leaves damaged by the tetranychus urticae is obtained by utilizing a hyperspectral imager, and the method is used for detecting and monitoring the tetranychus urticae. The rapid monitoring and identification method is developed by carrying out classification research on the hyperspectral reflectivity data of the plants, and optional technical support and scientific guarantee are provided for agricultural production.

Description

Technical Field

[0001] The present invention relates to the field of plant protection, in particular to the field of monitoring insect pests that harm plants, and in particular to an insect pest identification or monitoring method based on a hyperspectral spectrum. Background Art

[0002] The two-spotted spider mite, Tetranychus urticae, belongs to the genus Tetranychus of the family Tetranychidae and is a worldwide pest mite. It mainly gathers on the back of leaves to suck sap and feed. After the leaves are fed with sap, white spots will first appear on both sides of the main veins. As the damage worsens, the leaves will turn grayish white, inhibiting the normal photosynthesis. As a result, it causes serious yield reduction of field crops, fruit trees and vegetables (SUN Yue-hua, ZHI Jun-rui, TIAN Tian, ​​et al. The damage of Tetranychus urticae on physiological index of kidney bean [J]. Chinese Journal of Ecology, 2014, 31 (5): 843-854.). Among the resistant pests / mites that cause significant economic losses to crops worldwide, the two-spotted spider mite ranks fifth (Van Leeuwen T, Vontas J, Tsagkarakou A, Dermauw W, Tirry L. A caricide resistance mechanisms in the two-spotted spider mite Tetranychus urticae and other important Acari: a review. Insect Biochem Mol Biol. 2010, 40 (8): 563-572.).

[0003] The two-spotted spider mite is particularly difficult to control due to its small size, fast reproduction rate, short generation cycle, and ability to reproduce both sexually and parthenogenetically. Through early monitoring, effective control measures can be taken when the density is low, thereby reducing damage to plants, improving crop yield and quality, reducing control costs, and avoiding environmental pollution and resistance problems caused by the large-scale use of pesticides. In addition, early detection can also help formulate scientific and reasonable integrated control strategies to protect the health of agricultural ecosystems. Therefore, early detection of two-spotted spider mites is crucial because it helps to detect and control pests in a timely manner and prevent their large-scale reproduction and spread. It can be seen that the detection of pests is a core component of integrated pest management (Leybourne DJ, Musa N, Yang P. Can artificial intelligence be integrated into the best monitoring schemes to help achieve sustainable agriculture? Anentomological, management and computational perspective. Agricultural and Forest Entomology. 2024, 1–10.).

[0004] At present, there are three main types of monitoring methods for two-spotted spider mites. The first type is the visual monitoring method, which uses a magnifying glass or microscope to observe the damage of two-spotted spider mites. Although this method is simple and intuitive, it has many disadvantages, such as relying on the experience of the monitoring personnel, low efficiency, difficulty in early detection of two-spotted spider mites, being affected by environmental conditions, limited detection range and low accuracy. These shortcomings make it difficult to fully and accurately assess the actual damage of two-spotted spider mites, so it is usually necessary to combine other methods to improve the monitoring effect. The second type is the biochemical monitoring method, which uses PCR technology to detect two-spotted spider mite DNA in plant tissues, or detects changes in chlorophyll content in leaves, to achieve rapid and accurate monitoring and indirectly judge the damage of two-spotted spider mites. Although the biochemical method can provide valuable information in the monitoring of two-spotted spider mites, it also faces problems such as high cost, complex operation, high equipment requirements and possible environmental impact. The third type is the spectral information monitoring method, which uses spectral information to identify the characteristics of two-spotted spider mite infestation for monitoring, which improves monitoring efficiency and accuracy, reduces labor costs, and supports more scientific prevention and control decisions.

[0005] In recent years, hyperspectral has played an important role in pest and disease monitoring due to its rapid, non-destructive, water-resistant and micro-area resolution characteristics. Each pest has its own specific characteristic spectrum after damage. Based on the feasibility of using a portable hyperspectral instrument to distinguish between healthy leaves and leaves damaged by two-spotted spider mites, it can be used for early monitoring of two-spotted spider mites and prevention of spider mite outbreaks. Summary of the invention

[0006] In order to solve the above technical problems, this application uses hyperspectral imaging technology because of its fast, non-destructive, water-resistant and micro-area resolution. Each pest has its own specific characteristic spectrum after damage. A portable hyperspectral instrument is used to prove the feasibility of hyperspectral differentiation. A hyperspectral imaging camera is used to establish SVM and ResNet classification models, thereby developing a rapid monitoring and identification method to provide optional technical support and scientific guarantee for agricultural production.

[0007] To this end, this application proposes a sensitive, specialized, accurate and convenient identification method, which uses a portable hyperspectral imaging system to collect hyperspectral information before and after the two-spotted spider mite damages crop leaves, thereby achieving rapid monitoring and identification of the two-spotted spider mite damage to plant leaves.

[0008] This application first summarizes and compares the hyperspectral information of kidney bean leaves before and after the two-spotted spider mite damages by using a portable hyperspectral imaging camera, and finds that the characteristic peak intensity is different. Then, using a hyperspectral imaging camera and machine learning methods, the differences between leaves before and after the two-spotted spider mite damage are explored, and a rapid differentiation model for spider mite damage monitoring based on hyperspectral is established.

[0009] The present invention provides a method for establishing a hyperspectral-based spider mite damage identification or monitoring system, which is characterized by comprising the following steps:

[0010] (1) Pre-treat the plant samples;

[0011] (2) Collecting hyperspectral data of the sample to be tested;

[0012] (3) preprocessing the raw spectral data collected in step (2);

[0013] (4) Use machine learning to build a plant pest detection model;

[0014] Among them, the plant samples include leaves with different degrees of damage caused by spider mites, which are divided into clean leaves, leaves with slight damage, leaves with moderate damage, and leaves with severe damage; among them, leaves with uniform green color are regarded as healthy leaves, sporadic chlorotic white dot-like symptoms are regarded as slight damage, sporadic white dot-like damage symptoms gathered in blocks are regarded as moderate damage, and leaves with large-scale chlorosis and damage symptoms gathered in patches are regarded as severe damage, among which the traces of spider mites feeding on the leaves are the degree of chlorosis of the leaves.

[0015] In a specific embodiment, the spectrum acquisition includes the steps of preprocessing and data processing the hyperspectral data;

[0016] The data preprocessing steps described in A1 include MSE outlier removal, MSC radiation correction, and SG filtering operations;

[0017] The data processing steps described in A2 are: using dimensionality reduction to process the data, specifically using wavelength data with strong correlation with sample types: using principal component analysis PCA data dimensionality reduction technology and continuous projection algorithm SPA to reduce dimensionality.

[0018] Preferably, SVM and ResNet neural network modeling is performed on the full-band spectral data;

[0019] When establishing the SVM model, the optimal penalty coefficient c and kernel parameter g are selected by grid division and five-fold cross validation. Among them, c represents the preference weight of SVM for the two adjustment optimization directions of interval size and classification accuracy, that is, the tolerance to errors. The larger the c, the smaller the tolerance of the model to errors, and the easier it is to overfit. On the contrary, the model is more likely to underfit. If c is too large or too small, the generalization ability of the model will deteriorate.

[0020] The parameter g is a parameter of the RBF kernel function, which determines the distribution of the data after being mapped to the new feature space. The larger the g, the fewer the number of support vectors, and the faster the model training and prediction speeds.

[0021] When building a ResNet neural network model, the number of hidden layer nodes ranges according to the empirical formula Determine, where m is the number of input layer nodes, that is, the sample feature dimension; n is the number of output layer nodes, that is, the number of sample types; a is a real number between 1 and 10, and the h value with the highest prediction accuracy is selected as the number of hidden layer nodes of the model.

[0022] In a specific implementation manner, the accuracy of the training set and the test set of the constructed model can reach more than 90%, and the Kappa coefficient is more than 0.85.

[0023] Preferably, in the hyperspectral image preprocessing: MSE is used to evaluate the prediction accuracy of the model for spectral reflectance or radiation intensity, and MSE outlier removal is an iterative process for identifying and deleting data points in the data set that are significantly deviated from other data points due to measurement errors, data input errors or abnormal events.

[0024] The calculation formula of MSE is as follows:

[0025]

[0026] Where n is the number of spectral channels, Y i is the actual spectral data, is the mean spectrum;

[0027] MSC eliminates the scattering effect by converting the spectral data of the sample into the ratio of its mean spectrum to enhance the correlation between the spectrum and the data. The MSC calculation formula is as follows

[0028]

[0029] Among them, MSC i is the corrected spectrum, X i is the reflectance or radiation intensity of the ith band in the original spectrum, SC i is the scatter correction coefficient of the ith band, Mean(SC i ) is the average value of the scatter correction coefficients of all bands;

[0030] SG filtering is used to remove noise and mutations in the spectrum to improve the smoothness and readability of the data. It is a linear filtering method based on polynomial fitting, which smoothes the data by fitting a polynomial in a sliding window. SG filtering fits the data points with a low-order polynomial in a window of size 2m+1; specifically, assuming that a k-order polynomial is selected in each window:

[0031] y(i)=a0+a1x+a2x 2 +…+a k x k

[0032] Where x is the offset of each data point in the window relative to the center point;

[0033]

[0034] SG filtering finally finds a set of coefficients a0, a1, ..., a k The polynomial is made to best fit the original data points within the window by minimizing the following sum of squared errors.

[0035] More preferably, the hyperspectral image data classification method is as follows:

[0036] Data dimensionality reduction: The initial leaves are relatively healthy, and the principal component analysis (PCA) data dimensionality reduction technique is used. The original data is transformed into a new coordinate system through a linear transformation, so that the data variance on the first few coordinate axes in this new coordinate system is the largest, reducing the data dimension, removing noise, and discovering the most important structure in the data. After being infected to a moderate to severe degree, new peaks appear in the spectral characteristics, and the successive projections algorithm (SPA) is used for dimensionality reduction subsequently;

[0037] Furthermore, the successive projections algorithm (SPA) is used for feature band selection, and the steps are as follows: Denote the initial iteration vector as x k(0) , the number of variables extracted is N, and the spectral matrix is J columns:

[0038] Arbitrarily select 1 column (the j-th column) of the spectral matrix, and assign the j-th column of the modeling set to x j , denote x k(0) ;

[0039] Denote the set of positions of the column vectors not selected as s,

[0040]

[0041] Calculate the projections of x j on the remaining column vectors respectively:

[0042]

[0043] Extract the spectral wavelength of the maximum projection vector,

[0044] k(n)=arg(max(P(x j ))), j ∈ s)\*MERGEFORMAT(3)

[0045] Let x j =p x , j ∈ s0, n=n + 1. If n < N, then calculate cyclically according to formula (2);

[0046] The variables extracted are {x k(n) =0,..., N - 1};

[0047] For k(0) and N in each cycle, establish a multiple linear regression analysis (MLR) model respectively, and obtain the root mean square error of cross-validation (RMSECV) of the modeling set. For different candidate subsets, the k(0) and N corresponding to the smallest RMSECV are the optimal values. Generally, the fraction N of the characteristic wavelengths selected by SPA cannot be very large.

[0048] The present invention provides a hyperspectral-based spider mite damage identification or monitoring system obtained by the establishment method; optionally, it is implemented by computer programming; preferably, it is arranged as a local program or a remote processing program or arranged on a mobile device such as a mobile phone or a portable device.

[0049] Specifically, it includes the following modules:

[0050] A data entry module is used to enter leaf hyperspectral data of collected plant samples;

[0051] A data processing module, used to implement data processing in the system as claimed in claim 8;

[0052] Result output module, used to display the identification results;

[0053] Optionally, a data acquisition module is also included, which is used to collect the hyperspectral data of the leaves of the plant sample to be tested through a hyperspectral camera; specifically, the plant sample to be tested is placed in the center of the acquisition system, and the light source is turned on to observe the best imaging uniformity in the acquisition software. The scanning angle is determined for scanning, and the scanning time is to avoid damage to the sample by the light source.

[0054] The present invention provides a method for identifying or monitoring spider mite damage based on a hyperspectral method, which is characterized by comprising the following steps:

[0055] S1 Data collection: Take the leaves of the plants suspected of being or already infested by two-spotted spider mites to collect hyperspectral data;

[0056] S2 inputs the system as claimed in claim 8 or 9;

[0057] S3 obtains the detection or monitoring results, and obtains the early monitoring results of the two-spotted spider mite on the plant; that is, whether there is two-spotted spider mite damage, or the degree of damage by the two-spotted spider mite.

[0058] Beneficial effects of this application:

[0059] The two-spotted spider mite is particularly difficult to control due to its small size, fast reproduction rate, short generation cycle, and ability to reproduce both sexually and parthenogenetically. Therefore, early monitoring of the two-spotted spider mite is crucial, which will help to timely discover and control its occurrence and prevent its large-scale reproduction and spread. This application first uses a portable hyperspectral instrument to summarize and compare the hyperspectral data of the two-spotted spider mite before and after feeding and damage, and finds that there are differences in the hyperspectral data of the leaves after the two-spotted spider mite harms the plants, based on which they can be distinguished. That is, it verifies that the use of a portable hyperspectral instrument can be used to monitor the two-spotted spider mite before and after it harms plants. This is a step with practical application value and lays the foundation for the subsequent realization of instant field detection.

[0060] On this basis, using a hyperspectral instrument and machine learning methods, we explored the obvious differences between before and after the two-spotted spider mite feeding damage, and established a rapid differentiation model based on hyperspectral. Through unsupervised learning algorithms such as hyperspectral Kmeas clustering and spectral angle (SAM) clustering, different labels were given before and after the two-spotted spider mite feeding damage. Combined with supervised learning algorithms and models such as support vector machine (SVM) and ResNet neural network, the classification and identification of the degree of two-spotted spider mite feeding damage was achieved through comparative analysis of hyperspectral. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is the preprocessing process of hyperspectral images.

[0062] Figure 2 Schematic diagram of the portable hyperspectral imaging acquisition system.

[0063] Figure 3 This is the preprocessing of the raw hyperspectral data, which includes the raw spectral data, the data after outliers are removed by MSE, the data after multivariate scattering correction by MSC, and the data after SG filtering.

[0064] Figure 4 The difference between the spectral characteristics of healthy leaves and the spectral characteristics of insect-infested areas after pretreatment.

[0065] Figure 5 The original spectral curves of different spider mite infection times.

[0066] Figure 6 The difference in spectral curves of pest severity at different infection levels after pretreatment.

[0067] Figure 7 Schematic diagram of the decision hypersurface of SVM and the structure of ResNet.

[0068] Figure 8 Select feature distribution for the Successive Projections Algorithm (SPA).

[0069] Fig. 9 Spectral differentiation effect of healthy, lightly infected, moderately infected and severely infected leaves. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the embodiments of the present invention are further described in detail below in conjunction with the drawings of the embodiments of the present invention, but the described embodiments are only some embodiments of the present invention, not all. Based on the embodiments of the present invention, other embodiments of the present invention that are not creative work by technicians in this field belong to the scope of protection of the present invention.

[0071] Example 1: Preliminary detection of plant leaves damaged by spider mites based on a portable hyperspectral instrument

[0072] Experimental materials and instruments

[0073] The required kidney bean plants and two-spotted spider mites were bred and raised by the Predatory Mite Research Group of the Biological Control Laboratory, Institute of Plant Protection, Chinese Academy of Agricultural Sciences; the portable hyperspectrometer HC-6010 was from the Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Sciences.

[0074] The method for preliminary detection of plant leaves damaged by two-spotted spider mites based on a portable hyperspectral instrument is as follows:

[0075] (1) Prepare the test sample: prepare the sample leaves to be tested;

[0076] Five days after planting kidney beans, two-spotted spider mites were inoculated into clean leaves for continuous damage. Leaves with different degrees of damage were selected according to the damage time. A total of four degrees of damage were selected, including healthy leaves, lightly damaged leaves, moderately damaged leaves, and severely damaged leaves. Leaves that were uniformly green were considered healthy leaves, sporadic chlorotic white dot symptoms were considered light damage, sporadic white dot damage symptoms that appeared in blocks were considered moderate damage, and large areas of leaves turned green and the damage symptoms were concentrated in patches, considered severe damage.

[0077] (2) Find the damage and select characteristic plants

[0078] Take leaves with different degrees of damage by spider mites (healthy leaves, leaves with mild damage, leaves with moderate damage, leaves with severe damage) and select appropriate plants for hyperspectral data collection according to experimental needs.

[0079] (3) Spectral information collection

[0080] a. Hyperspectral system

[0081] The data is collected using a portable hyperspectral system, which includes a hyperspectral camera (1), a full-band light source (2), a scanning platform (3), and a support frame (4). The test plant (5) is placed inside. The structure is as follows: Figure 2 shown.

[0082] See Table 1 for parameter settings.

[0083] Table 1. Performance parameters of portable hyperspectral system.

[0084]

[0085]

[0086] b. Collection steps

[0087] Control the scanning platform and the position of the hyperspectral camera to ensure that the image plane of the hyperspectral camera is the clearest. Place the monitoring sample in the center of the acquisition system, turn on the light source and observe the best imaging uniformity in the acquisition software. Determine the scanning angle for scanning, and the scanning time should not be too long to avoid damage to the sample by the light source.

[0088] (4) Hyperspectral analysis

[0089] Hyperspectral analysis includes preprocessing and data processing steps on hyperspectral image data.

[0090] (a) Spectral data preprocessing

[0091] In the process of using hyperspectral image data, data preprocessing is a key step, including processing missing values, discovering outliers, etc. Specifically for hyperspectral images, each pixel contains 300 band intensity values, which are extremely sensitive to changes in the external environment. Therefore, it is necessary to correct and remove outliers to ensure the accuracy of data classification.

[0092] Hyperspectral image preprocessing process Figure 1 shown.

[0093] ① The mean square error (MSE) of hyperspectral data is an indicator used to measure the difference between the model prediction results and the actual observation values. For hyperspectral data, MSE can be used to evaluate the prediction accuracy of the model for spectral reflectance or radiation intensity.

[0094] The calculation formula of MSE is as follows:

[0095]

[0096] Where n is the number of spectral channels, Y i is the actual spectral data, is the mean spectrum. Generally, the mean of the spectrum is taken as the ideal value. By calculating the MSE, the square error between the actual spectrum data at each point and the actual observed value can be obtained. By retaining the data within a mean square error with the mean error, outliers can be eliminated.

[0097] ②Multiplicative Scatter Correction (MSC) is a commonly used spectral correction method. MSC can effectively eliminate the spectral differences caused by different scattering levels, thereby enhancing the correlation between the spectrum and the data. This method corrects the baseline shift and offset of the spectral data through the ideal spectrum. In practice, it is impossible to obtain the true ideal spectral data, so the average value of all spectral data is often assumed as the ideal spectrum. The MSC calculation formula is as follows

[0098]

[0099] Among them, MSC i is the corrected spectrum, X i is the reflectance or radiation intensity of the ith band in the original spectrum, SC i is the scatter correction coefficient of the ith band, Mean(SC i ) is the average value of the scatter correction coefficients for all bands.

[0100] ③SG filtering (Savitzky-Golay filtering) is a commonly used spectral smoothing method used to remove noise and mutations in the spectrum to improve the smoothness and readability of the data. SG filtering is a linear filtering method based on polynomial fitting, which smoothes the data by fitting a polynomial within a sliding window.

[0101] exist Figure 3 The changes in the spectral characteristic curves of the collected data after preprocessing are shown in Figure 2. It can be seen that the original spectral curve has a large error in the spectral peak due to inconsistent lighting conditions, while the spectral curves tend to be consistent after preprocessing.

[0102] (b) Hyperspectral image data classification

[0103] ①Data dimensionality reduction

[0104] The spectral data volume of hyperspectral images is large and there is redundant information, which will increase the computational complexity of the discriminant model and reduce the computational efficiency and accuracy of the model. Therefore, it is necessary to reduce the dimensionality of the spectral data and extract the wavelength data with strong correlation with the sample type. The variable iterative space shrinkage approach (VISSA), competitive adaptive reweight sampling method (CARS), iteratively retains informative variables (IRIV), and shuffled frog leaping algorithm (SFIA) are used to extract characteristic wavelengths, and the successive projections algorithm (SPA) is used to perform secondary dimension reduction on the characteristic wavelength spectral data. Different dimensionality reduction methods need to be used for different data. The principal component analysis PCA data dimensionality reduction technology is used for the initial healthy leaves. After the infection reaches moderate or severe levels, new peaks appear in the spectral characteristics, so the continuous projection algorithm SPA is used for dimensionality reduction.

[0105] ② Feature band selection

[0106] The successive projections algorithm (SPA) is a forward feature variable selection method. SPA uses the projection analysis of vectors. By projecting wavelengths onto other wavelengths and comparing the magnitudes of the projection vectors, the wavelength with the largest projection vector is selected as the candidate wavelength, and then the final characteristic wavelength is selected based on the calibration model. The variables selected by SPA are a combination with the least redundant information and the smallest collinearity. The brief steps of this algorithm are as follows.

[0107] Denote the initial iterative vector as x k(0) , the number of variables to be extracted is N, and the spectral matrix has J columns.

[0108] Arbitrarily select one column (the j-th column) of the spectral matrix, and assign the j-th column of the calibration set to x j , and denote x k(0) .

[0109] Denote the set of positions of the unselected column vectors as s,

[0110]

[0111] Calculate the projections of x j onto the remaining column vectors respectively:

[0112]

[0113] Extract the spectral wavelength with the largest projection vector,

[0114] k(n) = arg(max(P(x j ))), j ∈ s)\*MERGEFORMAT(3)

[0115] Let x j = p x , j ∈ s0, n = n + 1. If n < N, then calculate in a loop according to formula (2).

[0116] The extracted variables are {x k(n) = 0, …, N - 1}.

[0117] For k(0) and N in each loop, establish a multiple linear regression analysis (MLR) model respectively, and obtain the root mean square error of cross-validation (RMSECV) of the calibration set. For different candidate subsets, the k(0) and N corresponding to the smallest RMSECV are the optimal values. Generally, the fraction N of the characteristic wavelengths selected by SPA cannot be very large.

[0118] The specific results of the selected characteristic bands are shown in Figure 6 .

[0119] The criteria for detecting two-spotted spider mite damage are as follows:

[0120] (1) Model establishment

[0121] Support vector machine (SVM) is based on the principle of structural risk minimization in machine learning theory and can obtain the optimal solution in the case of small samples. The main idea of ​​SVM is to find the best classification surface that can distinguish positive examples from negative examples in the dimensional space. When facing nonlinear problems, the input space is transformed into a high-dimensional space with the help of kernel function, and then the best classification surface is obtained in this new space. Selecting an appropriate kernel function can improve the classification efficiency and make the division of samples clearer. The radial basis function (RBF) is selected as the kernel function.

[0122] ResNet (Residual Neural Network) is a deep convolutional neural network architecture used for data classification tasks. Compared with traditional deep networks, ResNet introduces residual connections, which directly add the output of the previous layer to the input of the subsequent layer through skip connections, thereby solving the gradient vanishing and model degradation problems in the deep network training process; at the same time, the residual connection enables the network to learn the residual function, thereby better capturing the mapping relationship between input and output. By stacking multiple residual blocks, ResNet can build very deep neural networks, such as ResNet-50, ResNet-101, etc. See the attached diagram for the SVM decision hypersurface and the Resnet network model. Figure 7 .

[0123] (2) Pest detection

[0124] SVM and ResNet neural network modeling were performed on the full-band spectral data, and the modeling results are shown in Table 2. When establishing the SVM model, the optimal penalty coefficient c and kernel parameter g were selected by grid division and five-fold cross-validation. c represents the preference weight of SVM for the two adjustment optimization directions of interval size and classification accuracy. That is, the larger the tolerance c for errors, the smaller the tolerance of the model for errors, and the easier it is to overfit; conversely, the model is more likely to underfit. c that is too large or too small will lead to poor generalization ability of the model. g is a parameter of the RBF kernel function, which determines the distribution of the data after mapping to the new feature space. The larger g is, the fewer the number of support vectors, and the faster the model training and prediction speeds. When establishing the ResNet neural network model, the number range of hidden layer nodes is based on the empirical formula Determine, where m is the number of input layer nodes, that is, the sample feature dimension; n is the number of output layer nodes, that is, the number of sample types; a is a real number between 1 and 10. The h value with the highest prediction accuracy is selected as the number of hidden layer nodes of the model. As shown in Table 2, the accuracy of the training set and test set of the constructed model can reach more than 90%, and the Kappa coefficient is above 0.85, which proves that the identification of spider mites through spectral information has a good effect.

[0125] The hyperspectral system monitored the spectral information of healthy leaves and leaves infested with spider mites for 10 days. The average spectrum was obtained by using Python through MSE outlier removal, MSC multivariate scatter correction and SG filtering. Figure 3 (Original spectral data, data after outliers are removed by MSE, data after multivariate scattering correction by MSC, and data after SG filtering, respectively).

[0126] By observing the spectral curve of the insect damage degree of leaves after pretreatment (such as Figure 4 As shown in the figure, it can be seen that in the wavelength range of 520-570nm, healthy leaves form a small reflection peak in the green band due to less absorption of chlorophyll, which is called the "green peak". In the wavelength range of 640-700nm, due to the strong absorption of chlorophyll, a red band absorption valley is formed, which is called the "red valley". Due to the porous thin-walled cell structure characteristics of the leaves, the near-infrared light is strongly reflected, forming the highest peak on the spectrum. When attacked by insect pests, the obvious changes in the spectrum are manifested as the gradual disappearance of the "green peak" and "red valley" in the visible light band, the increase in the slope of the "red edge", and the increase in the reflectivity of the near-infrared wavelength. The main reason for this phenomenon is that after the leaves are eaten by spider mites, the chlorophyll content is relatively reduced, the cell structure is destroyed, and the leaf spectrum moves to the infrared region, causing the appearance to transition from green to yellow. Therefore, it is feasible to use spectral differences to judge whether plant leaves are damaged by spider mites.

[0127] This example verifies the effectiveness of a portable hyperspectral instrument in distinguishing the spectra of healthy leaves from leaves infested by two-spotted spider mites, laying the foundation for achieving real-time monitoring in the field.

[0128] Example 2: Construction of a monitoring system for plant leaves damaged by spider mites based on a portable hyperspectral instrument

[0129] The portable system in Example 1 was used to monitor kidney bean leaves that had been infested for 3, 5, and 7 days, and the leaves were divided into lightly infected, moderately infected, and severely infected groups as experimental groups. In addition, a non-inoculated control group was set as a healthy group experiment, and the experimental group and the control group were tested simultaneously. The original spectral curves of the healthy group, lightly infected, moderately infected, and severely infected groups are shown in Figure 1. Figure 5 As shown in the figure, the hyperspectral data after MSE outlier removal, MSC multivariate scattering correction and SG filtering are Figure 6 shown.

[0130] After the acquisition is completed, the same hyperspectral image data preprocessing method as in Example 1 is used ( Figure 7 ). Note that in the multi-classification task of hyperspectral images, due to the redundancy of hyperspectral data, too much sample spectral data will inevitably lead to a large model size and difficulty in training, and it is not easy to find the relationship between spectral features. Therefore, it is necessary to extract features from the data, that is, to obtain the bands that are most sensitive to the spectral curve of the pest. After feature extraction using the SPA method, the top 20 most sensitive bands are selected as the basis for classification calculation. The distribution of feature points after band selection is shown in Figure 8 .

[0131] The training set (20,000 sample points) and the test set (1,000 sample points) were divided. The training set was divided into two groups, each with 10,000 sample points. SVM and ResNet models were used for classification training respectively, and the obtained training models were used to verify the test set, and the accuracy and Kappa index were calculated. The accuracy calculation formula is:

[0132]

[0133] Where S peedict-right Indicates the number of sample points with accurate prediction, S total Represents the total number of sample points in the test set.

[0134] The Kappa index is a statistic used to measure the agreement between two evaluators or classifiers on a classification problem. It takes into account that even if the evaluators randomly select the categories, a certain degree of agreement may occur. The value of the Kappa index ranges between -1 and 1, where 1 indicates perfect agreement, 0 indicates no agreement beyond random expectation, and negative values ​​indicate agreement below random expectation. Its calculation formula is:

[0135]

[0136] Among them, p0 calculates the sum of the number of samples correctly classified in all categories divided by the total number of samples, that is, the overall classification accuracy Accuracy, p e It is calculated based on the confusion matrix, taking into account the expected consistency of each class.

[0137] The classification results of different algorithms are shown in Table 2.

[0138] Table 2. Comparison of SVM and ResNet classification accuracy for full-band spectral data.

[0139]

[0140] It can be seen that both the SVM algorithm and the ResNet neural network model have high classification accuracy. The percentage of infected areas in the hyperspectral images of leaves with different degrees of spider mite infestation after classification is shown in the attached figure. Fig. 9 As shown in Table 3, principal component analysis (PCA) and linear discriminant analysis (LDA) were performed on the classification results to calculate the energy proportion of the original spectral components in the hyperspectral images of leaves with different infection degrees (i.e., the difference between the overall and healthy leaves as time and the degree of pest infection change), and the results are shown in Table 3. It can be seen that the infection trend of the leaves continues to deepen over time, and a reference model is established for the subsequent quantitative evaluation of the degree of infection.

[0141] Table 3. Distribution ratio of spectral energy in different infection degree areas using principal component analysis and discriminant linear analysis algorithms.

[0142] Infection severity PCA LDA healthy 0.7021 0.6942 Mild 0.5155 0.5201 Moderate 0.3256 0.4582 Severe 0.2025 0.3217

[0143] In summary, Example 1 verified the application potential of hyperspectral in monitoring plants infested by two-spotted spider mites. Example 2 successfully constructed a monitoring system for distinguishing plants infested by two-spotted spider mites based on a portable hyperspectral instrument by combining hyperspectral with SVM and ResNet neural networks, thus realizing early monitoring of two-spotted spider mites on plants.

Claims

1. A method for establishing a hyperspectral-based spider mite damage identification or monitoring system, characterized in that: The following steps are involved: (1) Pre-treat the plant samples; (2) Collecting hyperspectral data of the sample to be tested; (3) preprocessing the raw spectral data collected in step (2); (4) Use machine learning to build a plant pest detection model.

2. The method according to claim 1, characterized in that: The plant samples include leaves with different degrees of damage caused by spider mites, which are divided into clean leaves, leaves with slight damage, leaves with moderate damage, and leaves with severe damage; among them, leaves with uniform green color are regarded as healthy leaves, sporadic chlorotic white dot-like symptoms are regarded as slight damage, sporadic white dot-like damage symptoms gathered in blocks are regarded as moderate damage, and leaves with large-scale chlorosis and damage symptoms gathered in patches are regarded as severe damage, among which the traces of spider mites feeding on leaves are the degree of leaf chlorosis.

3. The establishment method according to claim 1, characterized in that: The spectrum acquisition includes the steps of preprocessing and data processing the hyperspectral data; The data preprocessing steps described in A1 include MSE outlier removal, MSC radiation correction, and SG filtering operations; The data processing steps described in A2 are: using dimensionality reduction to process the data, specifically using wavelength data with strong correlation with sample types: using principal component analysis PCA data dimensionality reduction technology and continuous projection algorithm SPA to reduce dimensionality.

4. The establishment method according to claim 2, characterized in that: Perform SVM and ResNet neural network modeling on full-band spectral data; When establishing the SVM model, the optimal penalty coefficient c and kernel parameter g are selected by grid division and five-fold cross validation. Among them, c represents the preference weight of SVM for the two adjustment optimization directions of interval size and classification accuracy, that is, the tolerance to errors. The larger the c, the smaller the tolerance of the model to errors, and the easier it is to overfit. On the contrary, the model is more likely to underfit. If c is too large or too small, the generalization ability of the model will deteriorate. The parameter g is a parameter of the RBF kernel function, which determines the distribution of the data after being mapped to the new feature space. The larger the g, the fewer the number of support vectors, and the faster the model training and prediction speeds. When building a ResNet neural network model, the number of hidden layer nodes ranges according to the empirical formula Determine, where m is the number of input layer nodes, that is, the sample feature dimension; n is the number of output layer nodes, that is, the number of sample types; a is a real number between 1 and 10, and the h value with the highest prediction accuracy is selected as the number of hidden layer nodes of the model.

5. The establishment method according to claim 4, characterized in that: The accuracy of the training set and test set of the constructed model can reach more than 90%, and the Kappa coefficient is above 0.

85.

6. The establishment method according to claim 3, characterized in that: In the hyperspectral image preprocessing, MSE is used to evaluate the prediction accuracy of the model for spectral reflectance or radiation intensity. MSE outlier removal is an iterative process used to identify and delete data points in the data set that are significantly deviated from other data points due to measurement errors, data input errors, or abnormal events. The calculation formula of MSE is as follows: Where n is the number of spectral channels, Y i is the actual spectral data, is the mean spectrum; MSC eliminates the scattering effect by converting the spectral data of the sample into the ratio of its mean spectrum to enhance the correlation between the spectrum and the data. The MSC calculation formula is as follows Among them, MSC i is the corrected spectrum, X i is the reflectance or radiation intensity of the ith band in the original spectrum, SC i is the scatter correction coefficient of the ith band, Mean(SC i ) is the average value of the scatter correction coefficients of all bands; SG filtering is used to remove noise and mutations in the spectrum to improve the smoothness and readability of the data. It is a linear filtering method based on polynomial fitting, which smoothes the data by fitting a polynomial in a sliding window. SG filtering fits the data points with a low-order polynomial in a window of size 2m+1; specifically, assuming that a k-order polynomial is selected in each window: y(i)=a0+a1x+a2x 2 +…+a k x k Where x is the offset of each data point in the window relative to the center point; SG filtering finally finds a set of coefficients a0, a1, ..., a k The polynomial is made to best fit the original data points within the window by minimizing the following sum of squared errors.

7. The establishment method according to claim 3, characterized in that: The hyperspectral image data classification method is as follows: Data dimensionality reduction: The leaves were relatively healthy at the beginning, so the principal component analysis (PCA) data dimensionality reduction technique was used to transform the original data into a new coordinate system through linear transformation, so that the data variance on the first few coordinate axes of this new coordinate system was maximized, reducing the dimension of the data, removing noise, and discovering the most important structure in the data; after the infection reached moderate or severe levels, new peaks appeared in the spectral characteristics, and the continuous projection algorithm (SPA) was used for dimensionality reduction later; Furthermore, the continuous projection algorithm SPA is used for feature band selection, and its steps are as follows: let the initial iteration vector be x k(0) , the number of extracted variables is N, and the spectral matrix is ​​j columns: Select any column (jth column) of the spectral matrix and assign the jth column of the modeling set to x j , remember x k(0) ; Let the set of unselected column vector positions be s, Calculate x separately j Projection onto the remaining column vectors: Extract the spectral wavelength of the maximum projection vector, k ( n ) arg ( max ( P ( x ) j )),j∈s)\*MERGER FORMAT(3) Let x j = p x , j ∈ s0, n = n + 1. If n < N, then calculate cyclically according to formula (2); The extracted variables are {x k(n) =0,…,N-1}; Corresponding to k(0) and N in each cycle, a multivariate linear regression analysis (MLR) model is established to obtain the root mean square error of interactive validation (RMSECV) of the modeling set, corresponding to different candidate subsets, where the k(0) and N corresponding to the smallest RMSECV are the optimal values. Generally, the characteristic wavelength fraction N selected by SPA cannot be too large.

8. A hyperspectral-based spider mite damage identification or monitoring system obtained according to the establishment method described in any one of claims 1 to 7; optionally, it is implemented by computer programming; preferably, it is arranged as a local program or a remote processing program or arranged on a mobile device such as a mobile phone or a portable device.

9. The system according to claim 8, characterized in that It includes the following modules: A data entry module is used to enter leaf hyperspectral data of collected plant samples; A data processing module, used to implement data processing in the system as claimed in claim 8; Result output module, used to display the identification results; Optionally, a data acquisition module is also included, which is used to collect the hyperspectral data of the leaves of the plant sample to be tested through a hyperspectral camera; specifically, the plant sample to be tested is placed in the center of the acquisition system, and the light source is turned on to observe the best imaging uniformity in the acquisition software. The scanning angle is determined for scanning, and the scanning time is to avoid damage to the sample by the light source.

10. A method for identifying or monitoring spider mite damage based on hyperspectral, characterized in that: The steps include: S1 Data collection: Take the leaves of the plants suspected of being or already infested by two-spotted spider mites to collect hyperspectral data; S2 inputs the system as claimed in claim 8 or 9; S3 obtains the detection or monitoring results, and obtains the early monitoring results of the two-spotted spider mite on the plant; that is, whether there is two-spotted spider mite damage, or the degree of damage by the two-spotted spider mite.

Images