A high-precision plant spectral detection method based on data fusion

Abstract

The present invention discloses a high-precision plant spectrum detection method based on data fusion, which belongs to the technical field of disease identification. The present invention effectively compensates the pixel values ​​of the first reflection thermal image, the first refraction thermal image, the second reflection thermal image, and the second refraction thermal image through environmental parameters, reduces measurement errors, and improves the accuracy of data. Furthermore, the disease probability of plant slices is predicted based on the first reflection thermal image, the first refraction thermal image, the second reflection thermal image, the second refraction thermal image, Raman spectrum data, and infrared spectrum data, and the preliminary detection is completed to exclude healthy plant slices, thereby speeding up the detection speed and avoiding congestion caused by the spectrum data of multiple groups of plant slice samples waiting to be processed during the production process of the assembly line. Finally, a disease classification model is generated through Raman spectrum data and infrared spectrum data, and the two types of data complement each other, thereby improving the accuracy and sensitivity of disease detection.

Description

Technical Field

[0001] The present invention relates to the technical field of plant internal disease recognition, and in particular to a high-precision plant spectrum detection method based on data fusion. Background Art

[0002] When a plant is diseased, in order to resist fungal and bacterial infection, the content of biological macromolecules such as chlorophyll, protein, phenolic compounds, and hormone substances will not be within the normal range. Obtaining information on the chemical composition and structure of the plant through the spectrum is very useful for detecting subtle chemical changes caused by diseases. The Chinese patent application with publication number CN115753738A discloses a pathological diagnosis method based on a biomarker enhanced Raman spectroscopy database. The method establishes a database based on the theoretical Raman spectra of biomarkers and the enhanced Raman spectra of measured biomarker proteins and tissue sections, and uses a deep learning model to judge the degree of pathological expression of the tissue to be tested. The Raman spectrum will have overlapping characteristic peaks. The above method establishes a database through the theoretical Raman spectra of biomarkers and the enhanced Raman spectra of measured biomarker proteins and tissue sections, and the deep learning model generated by training is not accurate.

[0003] The Chinese patent application with publication number CN111458309A discloses a qualitative method for vegetable oil based on near infrared-Raman analysis. This method measures the Raman spectrum and near infrared spectrum of the same sample, compares the infrared-Raman analysis characteristic graph of the sample to be tested with the standard spectrum of the known source sample, and evaluates the confidence level of the conformity between the sample to be tested and the standard spectrum to know the type of the sample to be tested. Multispectral data fusion technology can significantly improve the detection accuracy, but due to the large amount of calculation and long time consumption of data fusion modeling, this patented technology is mainly used for laboratory measurement. In order to adapt to industrial production, it is necessary to propose a fast and high-precision plant spectral detection method. The existing technology needs to be further improved. Summary of the invention

[0004] In order to solve the defects of the above-mentioned prior art, the present invention proposes a high-precision plant spectral detection method based on data fusion. The present invention collects thermal imaging data, Raman spectral data and infrared spectral data of plant slices, first obtains key parameters through preliminary data processing, predicts the disease status of plants, and then marks plant slices that may have diseases, and judges the diseases of plant slices according to the disease classification model. The present invention can significantly improve the detection efficiency and avoid congestion caused by the spectral data of multiple groups of plant slice samples waiting to be processed during the production line.

[0005] The technical solution of the present invention is achieved in this way:

[0006] A high-precision plant spectrum detection method based on data fusion, comprising the following steps:

[0007] Step 1: prepare multiple groups of plant slices, the plant slices enter the first detection channel in turn, the first light generator emits a first laser to the plant slices, the first light receiver collects Raman spectrum data, and the second light receiver collects a first reflection thermal image and a first refraction thermal image;

[0008] Step 2: extracting corresponding characteristic peaks in the Raman spectrum data according to the disease markers, and calculating a first content set of the disease markers;

[0009] Step 3: After a preset time, the plant slice enters the second detection channel, and the third optical receiver collects the second reflection thermal image and the second refraction thermal image;

[0010] Step 4: adjusting the first reflection thermal image, the first refraction thermal image, the second reflection thermal image, and the second refraction thermal image according to the environmental parameters, and then generating the maximum temperature rise parameter, the maximum temperature difference parameter, and the temperature attenuation parameter;

[0011] Step 5: The plant slice enters the third detection channel, the second light generator emits a second laser to the plant slice, and the fourth light receiver collects infrared spectrum data;

[0012] Step 6: searching for corresponding characteristic absorption peaks in the infrared spectrum data according to the disease markers, and calculating a second content set of the disease markers;

[0013] Step 7: predicting the disease probability of the plant slice according to the first content set, the second content set, the maximum temperature rise parameter, the temperature difference parameter and the temperature attenuation parameter of the disease markers; if the disease probability is greater than the probability threshold, marking the plant slice and proceeding to step 8; otherwise, returning to step 1;

[0014] Step 8: Create a disease classification model based on the Raman spectrum and infrared spectrum of the plant sample, input the Raman spectrum data and infrared spectrum data of the plant slice into the disease classification model, and predict the disease of the plant slice.

[0015] In the present invention, the plant slices are citrus fruit slices, the wavelength of the first laser is 785 nm, and the wavelength of the second laser is 10 μm.

[0016] In the present invention, in step 2, the disease markers include chlorophyll, protein, chlorogenic acid, jasmonic acid, chitin, and phenylalanine.

[0017] In the present invention, in step 4, the environmental parameters include atmospheric temperature T, atmospheric pressure P, relative humidity H, observation distance d, and meteorological visibility V. The atmospheric transmittance η is calculated according to the environmental parameters, and the first reflection thermal image, the first refraction thermal image, the second reflection thermal image, and the second refraction thermal image are adjusted according to the atmospheric transmittance. , a1 is the absorption coefficient, a 2 is the scattering coefficient, , , b 1 is the first constant, b 2 is the second constant, R 1 is the specific gas constant of water vapor, T 0 is the reference temperature, P 0 is the reference air pressure, λ is the wavelength of plant thermal radiation, c 1xy =ηc' 1xy , c' 1xy is the pixel value of the first reflection thermal image at the pixel coordinate (x, y), c 1xy is the pixel value of the adjusted first reflected thermal image at the pixel coordinate (x, y).

[0018] In the present invention, in step 4, the pixel value of each pixel coordinate in the adjusted first reflection thermal image, the first refraction thermal image, the second reflection thermal image, and the second refraction thermal image is converted into a temperature value, and the maximum temperature rise ΔT 1 =max{T 1xy -T 2xy}, maximum temperature difference ΔT 2 =max{|T 1xy -T 2xy |}, T 1xy is the temperature value of the adjusted first reflected thermal image at the pixel coordinate (x, y), T 2xy is the temperature value of the adjusted first refraction thermal image at the pixel coordinate (x, y), the maximum temperature attenuation rate , T 3xy is the temperature value of the adjusted second reflection thermal image at the pixel coordinate (x, y), T 4xy is the temperature value of the adjusted second refraction thermal image at the pixel coordinate (x, y), and t is the preset time length.

[0019] In the present invention, in step 7, the first content set = {q 11 ,q 12 ,q 13 ,q 14 ,q 15 ,q 16}, the second content set = {q 21 ,q 22 ,q 23 ,q 24 ,q 25 ,q 26}, q 11 is the chlorophyll content obtained from Raman spectroscopy data, q 12 is the protein content obtained from Raman spectroscopy data, q 13The chlorogenic acid content obtained from Raman spectroscopy data, q 14 The jasmonic acid content obtained from Raman spectroscopy data, q 15 The chitin content obtained from Raman spectroscopy data, q 16 The phenylalanine content obtained from Raman spectroscopy data, q 21 The chlorophyll content obtained from infrared spectroscopy data, q 22 The protein content obtained from infrared spectroscopy data, q 23 The chlorogenic acid content obtained from infrared spectroscopy data, q 24 The jasmonic acid content obtained from infrared spectroscopy data, q 25 The chitin content obtained from infrared spectroscopy data, q 26 The phenylalanine content obtained from infrared spectroscopy data.

[0020] In the present invention, in step 7, the light absorption rate is calculated based on the maximum temperature rise, the thermal conductivity is calculated based on the maximum temperature difference, and the thermal diffusivity is calculated based on the maximum temperature decay. The light absorption rate , b 3 is the specific heat capacity of the citrus fruit slice, M is the mass of the citrus fruit slice, Δt is the illumination time, P is the incident light power of the first light generator, and the thermal conductivity , ρ is the heat flux density of the citrus fruit slice, D is the thickness of the citrus fruit slice, and the thermal diffusivity .

[0021] In the present invention, the disease probability , q 1 is the reference content of chlorophyll, q 2 is the reference content of protein, q 3 is the reference content of chlorogenic acid, q 4 is the reference content of jasmonic acid, q 5 is the reference content of chitin, q 6 is the reference content of phenylalanine, w 1 is the first weight, w 2 is the second weight, w 3 is the third weight, w 4 is the fourth weight, γ 0 is the reference value of the light absorption rate, δ 0 is the reference value of the thermal conductivity, σ 0 is the reference value of the thermal diffusivity.

[0022] In the present invention, in step 8, Raman spectra and infrared spectra of multiple groups of plant samples are collected, the disease type and disease area are marked on each Raman spectrum and infrared spectrum, the marked Raman spectrum and infrared spectrum are divided into a training set and a validation set, the disease classification model is trained using a supervised deep learning method through the training set, the training error is calculated according to the loss function, and then at least one parameter of the disease classification model is adjusted according to the training error, and finally the disease classification model is verified and output using the validation set.

[0023] The high-precision plant spectrum detection method based on data fusion implemented in the present invention has the following beneficial effects: the present invention calculates the atmospheric transmittance through environmental parameters, and effectively compensates the pixel values ​​of the first reflection thermal image, the first refraction thermal image, the second reflection thermal image, and the second refraction thermal image based on the atmospheric transmittance, thereby reducing the measurement error caused by atmospheric absorption and scattering, and improving the accuracy of the data. Further, the maximum temperature rise parameter, the temperature difference parameter, and the temperature attenuation parameter are calculated according to the first reflection thermal image, the first refraction thermal image, the second reflection thermal image, and the second refraction thermal image, and the corresponding characteristic peaks in the Raman spectrum data and the infrared spectrum data are extracted according to the disease marker, and the first content set and the second content set of the disease marker are calculated. The disease probability of the plant slice is predicted by the first content set, the second content set, the maximum temperature rise parameter, the temperature difference parameter, and the temperature attenuation parameter of the disease marker, and the preliminary detection is completed to exclude healthy plant slices, simplify the detection steps, speed up the detection speed, and avoid congestion caused by the spectral data of multiple groups of plant slice samples waiting for processing during the production line. Finally, a disease classification model is generated through Raman spectral data and infrared spectral data. The two types of data complement each other and improve the accuracy and sensitivity of disease detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a flow chart of the high-precision plant spectrum detection method based on data fusion of the present invention;

[0025] Figure 2 It is a schematic diagram of the equipment arrangement of the present invention;

[0026] Figure 3 It is a schematic diagram of plant slices of the present invention;

[0027] Figure 4 is a schematic diagram of the first detection channel of the present invention;

[0028] Figure 5 is a schematic diagram of Raman spectroscopy data of the present invention;

[0029] Figure 6 is a schematic diagram of the second detection channel of the present invention;

[0030] Figure 7 is a schematic diagram of the third detection channel of the present invention;

[0031] Figure 8 Schematic diagram of light absorption rate of bacterial decay with different defect areas in the present invention;

[0032] Fig. 9 It is a schematic diagram of the heat diffusion rate of soft rot of different areas in the present invention. DETAILED DESCRIPTION

[0033] In order to more clearly understand the purpose, technical solutions and advantages of the present application, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments.

[0034] When a disease occurs inside a plant, it may cause a temperature change. This change can be either an increase or a decrease in temperature, depending on the nature and location of the disease and how the plant responds to the disease. In the early stages of infection, the disease characteristics are not obvious and the disease cannot be identified based on the plant's appearance image. The plant's light absorption rate, thermal conductivity, and thermal diffusion rate are measured to further provide feedback on the plant's disease.

[0035] Plant disease markers are biological or chemical indicators that can be used to identify and monitor whether plants are infected by pathogens or affected by environmental stress. These disease markers can help detect diseases. Raman spectroscopy and Fourier transform attenuated total reflectance infrared spectroscopy can detect the presence of diseases before the appearance of plant tissues changes.

[0036] Fourier transform attenuated total reflection infrared spectroscopy mainly detects the stretching vibration mode of the internal bonds of molecules, while Raman spectroscopy mainly reflects the change of molecular polarizability. The two spectra may produce different characteristic peaks for the same molecule, providing complementary information. Fourier transform attenuated total reflection infrared spectroscopy is suitable for the study of whole plant tissues or large plant cell populations, large-scale screening and rapid diagnosis. It can directly process complex matrices and can efficiently obtain a large amount of chemical information. On the other hand, Raman spectroscopy is suitable for exploring the disease microenvironment and heterogeneity at the single cell level due to its high spatial resolution and non-destructive characteristics. The combination of Fourier transform attenuated total reflection infrared spectroscopy and Raman spectroscopy can more accurately identify diseases. Embodiment 1

[0037] like Figures 1 to 7 As shown, the high-precision plant spectral detection method based on data fusion of the present invention includes the following steps.

[0038] Step 1: Prepare multiple groups of plant slices, and the plant slices enter the first detection channel in turn. The first light generator emits a first laser to the plant slices, the first light receiver collects Raman spectrum data, and the second light receiver collects the first reflection thermal image and the first refraction thermal image. Figure 2As shown, multiple groups of plant slices are placed at the starting station at a certain time interval, and the conveyor belt drives the plant slices into the first detection channel. Figure 3 As shown, the plant slices are citrus fruit slices, and the plant slices are placed at the starting position. Figure 4 As shown, the first detection channel includes a first light generator, a first light receiver and two groups of second light receivers. The first light generator emits a first laser to the plant slice, which is collected by the first light receiver and the two groups of second light receivers. One group of second light receivers is located on the reflection surface of the first laser, and the other group of second light receivers is located on the refraction surface of the first laser. The wavelength of the first laser is 785nm.

[0039] Step 2: Extract the corresponding characteristic peaks in the Raman spectrum data according to the disease markers, and calculate the first content set of the disease markers. Disease markers include chlorophyll, protein, chlorogenic acid, jasmonic acid, chitin, and phenylalanine. Preprocess the Raman spectrum data to reduce noise, then find the corresponding characteristic peaks in the Raman spectrum data according to the disease markers, and calculate the first content set of the disease markers. The preprocessing includes removing cosmic rays, baseline correction, noise smoothing, data clipping, peak alignment, normalization, and derivative processing. Figure 5 As shown, the horizontal axis of the Raman spectrum data is the wave number, and the unit is cm -1 , reflects the energy change of molecular vibration, the ordinate is the light intensity, the unit is W / m², which represents the Raman scattering intensity at the corresponding wave number. The noise caused by fluorescence spectrum and experimental conditions in Raman spectroscopy data is eliminated through preprocessing.

[0040] In the Raman spectrum data, the characteristic peak of chlorophyll is 750 cm -1 、980cm -1 Up to 1000cm -1 、1050cm -1 Up to 1070cm -1 、1290cm -1 、1310cm -1 、1510cm -1 Up to 1520cm -1 、1650cm -1 The characteristic peak of protein is 830 cm -1 、850cm -1 、1004cm -1 、1230cm -1 Up to 1300cm -1 、1450cm -1 、1650cm -1 The characteristic peak of chlorogenic acid is 830 cm -1 、1080cm -1 、1170cm -1、1270cm -1 、1510cm -1 、1600cm -1 The characteristic peak of jasmonic acid is 800 cm -1 、980cm -1 、1250cm -1 、1440cm -1 、1650cm -1 、1740cm -1 The characteristic peak of chitin is 1070 cm -1 、1155cm -1 、1315cm -1 、1425cm -1 、1650cm -1 、2870cm -1 、2920cm -1 The characteristic peak of phenylalanine is 1004 cm -1 、1200cm -1 Up to 1300cm -1 、1445cm -1 、1598cm -1 、1605cm -1 、3060cm -1 .

[0041] In the Raman spectrum data, the characteristic peaks of protein and chlorogenic acid are at 980 cm -1 Overlapping, the characteristic peaks of chlorophyll, jasmonic acid and chitin are at 1650cm -1 Overlap, separate the characteristic peaks of different disease markers through spectral deconvolution technology. Because the Raman signal intensity (peak height or peak area) and the concentration of the corresponding disease marker are linearly or nonlinearly related within a certain range, a multivariate regression model of spectral intensity and concentration is established, and the disease markers are quantified through the partial least squares regression algorithm, and then the first content set of disease markers is obtained. In addition to measuring disease markers through Raman spectral data, the Raman spectral intensity of disease markers can also be significantly enhanced through surface enhanced Raman scattering technology, so that low-concentration substances can also be accurately detected.

[0042] Step 3: After a preset time, the plant slice enters the second detection channel, and the third optical receiver collects the second reflection thermal image and the second refraction thermal image. Figure 2 As shown, the conveyor belt drives the plant slices from the first detection channel to the second detection channel after a preset time. During the preset time, the plant slices are isolated from the external environment by the black box cover to prevent the external light from affecting the temperature of the plant slices, resulting in inaccurate second reflection thermal images and second refraction thermal images collected later. Figure 6As shown, the second detection channel includes two groups of third light receivers, which receive the infrared thermal radiation emitted by the plant slices to form a second reflection thermal image and a second refraction thermal image. In this embodiment, the preset time length is 10s, and the speed of the conveyor belt is adjusted by setting the preset time length.

[0043] Step 4: Adjust the first reflection thermal image, the first refraction thermal image, the second reflection thermal image, and the second refraction thermal image according to the environmental parameters, and regenerate the maximum temperature rise parameter, the maximum temperature difference parameter, and the temperature attenuation parameter. The environmental parameter set includes atmospheric temperature T, atmospheric pressure P, relative humidity H, observation distance d, and meteorological visibility V. The unit of T is ℃, the unit of P is hPa, H has no unit, the unit of d is m, and the unit of V is km. Calculate the atmospheric transmittance η according to the environmental parameters, and adjust the first reflection thermal image, the first refraction thermal image, the second reflection thermal image, and the second refraction thermal image according to the atmospheric transmittance to improve the accuracy of the data. The calculation of atmospheric transmittance needs to comprehensively consider the effects of gas absorption and particle scattering. , a 1 is the absorption coefficient, a 2 is the scattering coefficient. , In the calculation formula of the absorption coefficient, the unit of 6.112 is hPa, the unit of 243.5 is ℃, the unit of 273.15 is ℃, and the unit of 550 in the calculation formula of the scattering coefficient is nm, and 3.912 has no unit. b 1 is the first constant, in m 2 / kg,b 2 is the second constant, unitless. The values ​​of the first constant and the second constant are related to the wavelength and experimental conditions. Since the wavelength of plant thermal radiation is in the infrared light band, in this embodiment, b 1 =0.05m 2 / kg,b 2 =0.5. 1 is the specific gas constant of water vapor. In this embodiment, R 1 =461.5J / (kg·K) T 0 is the reference temperature, P 0 is the reference pressure. In this embodiment, T 0 =288.15K=15℃, P 0 =1013.25hPa. λ is the wavelength of plant thermal radiation, in nm. In this embodiment, λ=50nm.

[0044] c 1xy =ηc' 1xy , c' 1xy is the pixel value of the first reflection thermal image at the pixel coordinate (x, y), c 1xyis the pixel value of the adjusted first reflection thermal image at the pixel coordinate (x, y). 2xy =ηc' 2xy , c' 2xy is the pixel value of the first refraction thermal image at the pixel coordinate (x, y), c 2xy is the pixel value of the adjusted first refraction thermal image at pixel coordinate (x, y). 3xy =ηc' 3xy , c' 3xy is the pixel value of the second reflection thermal image at the pixel coordinate (x, y), c 3xy is the pixel value of the adjusted first reflection thermal image at the pixel coordinate (x, y). 4xy =ηc' 4xy , c' 4xy is the pixel value of the second refraction thermal image at the pixel coordinate (x, y), c 4xy is the pixel value of the adjusted second refraction thermal image at the pixel coordinate (x, y).

[0045] The pixel value of each pixel coordinate in the adjusted first reflection thermal image, the first refraction thermal image, the second reflection thermal image, and the second refraction thermal image is converted into a temperature value, and the maximum temperature rise ΔT 1 =max{T 1xy -T 2xy}, maximum temperature difference ΔT 2 =max{|T 1xy -T 2xy |}, T 1xy is the temperature value of the adjusted first reflected thermal image at the pixel coordinate (x, y), T 2xy is the temperature value of the adjusted first refraction thermal image at the pixel coordinate (x, y), the maximum temperature attenuation rate , T 3xy is the temperature value of the adjusted second reflection thermal image at the pixel coordinate (x, y), T 4xy is the temperature value of the adjusted second refraction thermal image at the pixel coordinate (x, y), and t is the preset time length.

[0046] Step 5: The plant slice enters the third detection channel, the second light generator emits the second laser to the plant slice, and the third light receiver collects infrared spectrum data. Figure 7 As shown, the third detection channel includes a second light generator and a fourth light receiver. The second light generator emits a second laser to the plant slice, which is collected by the fourth light receiver. The wavelength of the second laser is 10 μm.

[0047] Step 6: According to the disease marker, the corresponding characteristic absorption peak in the infrared spectrum data is searched, and the second content set of the disease marker is calculated. The infrared spectrum data is Fourier transform attenuated total reflection infrared spectrum data, and the infrared spectrum data is preprocessed to reduce the noise of the infrared spectrum data. The preprocessing of the infrared spectrum data includes baseline correction, noise smoothing, normalization, data interception, differential spectrum, derivative processing, atmospheric interference removal, calibration and verification.

[0048] In infrared spectrum data, the characteristic peak of chlorophyll is 980cm -1 Up to 1000cm -1 、1050cm -1 Up to 1070cm -1 、1290cm -1 、1310cm -1 、1650cm -1 The characteristic peak of protein is 1240 cm -1 、1540cm -1 、1650cm -1 The characteristic peak of chlorogenic acid is 830 cm -1 、1080cm -1 、1270cm -1 、1610cm -1 、1740cm -1 The characteristic peak of jasmonic acid is 800 cm -1 、980cm -1 、1250cm -1 、1440cm -1 、1650cm -1 、1740cm -1 The characteristic peak of chitin is 1070 cm -1 、1155cm -1 、1315cm -1 、1425cm -1 、1650cm -1 、2870cm -1 、2920cm -1 The characteristic peak of phenylalanine is 1004 cm -1 、1200cm -1 Up to 1300cm -1 、1445cm -1 、1598cm -1 、1605cm -1 、3060cm -1First, the overlapping characteristic peaks of different disease markers in the infrared spectrum data are separated by second-order derivative spectroscopy or deconvolution technology, and then the disease markers are quantified by partial least squares discriminant analysis, thereby obtaining a second content set of disease markers.

[0049] Step 7: Predict the disease probability of the plant slice according to the first content set, the second content set, the maximum temperature rise parameter, the temperature difference parameter and the temperature attenuation parameter of the disease marker. If the disease probability is greater than the probability threshold, mark the plant slice and proceed to step 8, otherwise return to step 1. The first content set = {q 11 ,q 12 ,q 13 ,q 14 ,q 15 ,q 16}, the second content set = {q 21 ,q 22 ,q 23 ,q 24 ,q 25 ,q 26}, q 11 is the chlorophyll content obtained from Raman spectroscopy data, q 12 is the protein content obtained from Raman spectroscopy data, q 13 is the chlorogenic acid content obtained from Raman spectroscopy data, q 14 is the jasmonic acid content obtained from Raman spectroscopy data, q 15 is the chitin content obtained from Raman spectroscopy data, q 16 is the phenylalanine content obtained from Raman spectroscopy data, q 21 is the chlorophyll content obtained from infrared spectroscopy data, q 22 is the protein content obtained from infrared spectroscopy data, q 23 is the chlorogenic acid content obtained from infrared spectroscopy data, q 24 is the jasmonic acid content obtained from infrared spectroscopy data, q 25 is the chitin content obtained from infrared spectroscopy data, q 26 is the phenylalanine content obtained from infrared spectrum data. The method for generating the disease probability is described in detail in Example 2.

[0050] Step 8: Create a disease classification model based on the Raman spectrum and infrared spectrum of the plant sample, input the Raman spectrum data and infrared spectrum data of the plant slice into the disease classification model, and predict the disease of the plant slice. Collect Raman spectra and infrared spectra of multiple groups of plant samples, mark the disease type and disease area on each Raman spectrum and infrared spectrum, divide the marked Raman spectrum and infrared spectrum into a training set and a validation set, train the disease classification model with a supervised deep learning method through the training set, calculate the training error according to the loss function, and then adjust at least one parameter of the disease classification model according to the training error, and finally use the validation set to verify and output the disease classification model.

[0051] In addition, the disease classification model can also be generated by the following method: the Raman spectrum disease type and disease area label is used as the first label, and the infrared spectrum disease type and disease area label is used as the second label. First, multiple sample Raman spectrum data are divided into a test set and a validation set, and multiple basic models are selected. Using a cross-validation method, each basic model is trained through the test set to obtain a first recognition result. The accuracy of each basic model is determined based on the first label and the first recognition result, and the basic model with the highest accuracy is selected as the target basic model.

[0052] Then, the target basic model is used to predict the test set to obtain the first identification label. Multiple comprehensive models are selected, the parameters of the comprehensive models are optimized, and each comprehensive model is trained with the first identification label. The second identification result is predicted by the comprehensive model for the validation set, and the accuracy of each comprehensive model is obtained by comparing the second identification result with the first label, and the comprehensive model with the highest accuracy is selected as the target comprehensive model. The target basic model and the target comprehensive model are used as the first classification model of the Raman spectral data.

[0053] Finally, the second classification model of infrared spectrum data is determined according to the above method, and the disease classification model is generated according to the first classification model and the second classification model. The method for generating the disease identifier is, for example, Bayesian discriminant analysis. The discriminant function is constructed according to the Bayesian discriminant analysis method, and the recognition results obtained by the first classification model and the recognition results obtained by the second classification model are used to generate the final detection result through simple averaging, weighted averaging or multiplication rules. The Raman spectrum data and the infrared spectrum data complement each other, thereby enhancing the accuracy of disease classification.

[0054] The basic model is, for example, a support vector machine, a random forest, a nearest neighbor algorithm, a neural network, etc., and the comprehensive model is, for example, a logistic regression, a random forest, a machine learning, etc. The method for generating the first classification model divides the Raman spectral data of different spectral ranges into multiple data sets and cooperates with different models to achieve the purpose of identifying diseases, making full use of all spectral information of the Raman spectral data, and improving the classification accuracy and stability of the first classification model. Embodiment 2

[0055] like Figures 8 to 9 As shown, this embodiment further discloses a preferred method for generating disease probability.

[0056] First, the light absorption rate is calculated according to the maximum temperature rise, the thermal conductivity is calculated according to the maximum temperature difference, and the thermal diffusion rate is calculated according to the maximum temperature attenuation. Then the first weight, second weight, third weight, and fourth weight are set, and finally the above values ​​are substituted into the calculation formula of the disease probability.

[0057] Light Absorption , b 3 is the specific heat capacity of the citrus fruit slice, in J / (kg·K), M is the mass of the citrus fruit slice, in kg, Δt is the illumination time, in s, P is the incident light power of the first light generator, in W, in this embodiment, b 3 =4.0×10 3 J / (kg·K), M=0.015kg, Δt=3s. Thermal conductivity , ρ is the heat flux density of the citrus fruit slice, in W / m², D is the thickness of the citrus fruit slice, in m. In this embodiment, ρ=-50W / m², D=5×10 -3 m. Substitute the thermal conductivity, specific heat capacity of citrus fruit slices, maximum temperature decay, and heat flux density of citrus fruit slices into the calculation formula of thermal diffusion rate to obtain the thermal diffusion rate. .

[0058] like Figure 8 As mentioned above, the horizontal axis is the diseased area, and the vertical axis is the light absorption rate. When the diseased area exceeds a certain range, as the diseased area increases, the higher the degree of bacterial decay, the greater the light absorption rate. The light absorption rate of healthy citrus tissue in the visible light range is low, roughly ranging from 10% to 20%. Therefore, the light absorption rate of the plant slice can be used to determine whether the plant slice is diseased. Fig. 9 As mentioned above, the horizontal axis is the diseased area and the vertical axis is the heat diffusion rate. When the diseased area exceeds a certain range, as the diseased area increases, the higher the degree of soft rot, the lower the heat diffusion rate. Because the thermal conductivity and density of diseased plant tissues are usually lower than those of healthy plant tissues, as the degree of decay increases, the tissue structure and chemical composition change more, which reduces the heat diffusion rate. The heat diffusion rate of healthy citrus tissue is 1.0×10 -7 m² / s to 2.0×10 -7m² / s, and the heat diffusion rate decreases significantly with the increase of disease area and degree of decay. In addition, the thermal conductivity of healthy citrus tissue is between 0.2W / (m·K) and 0.4W / (m·K). In citrus tissue with callus, heat production gradually decreases, resulting in a decrease in thermal conductivity due to cell fluid leakage.

[0059] According to the above analysis, the probability of plant disease is related to light absorption rate, thermal conductivity, and thermal diffusion rate. When calculating the probability of disease, only the values ​​of light absorption rate, thermal conductivity, and thermal diffusion rate are substituted without actual units. ,q 1 is the reference content of chlorophyll, q 2 is the protein content, q 3 is the standard content of chlorogenic acid, q 4 is the standard content of jasmonic acid, q 5 is the base content of chitin, q 6 is the standard content of phenylalanine, w 1 is the first weight, w 2 is the second weight, w 3 is the third weight, w 4 is the fourth weight, γ 0 is the reference value of light absorption rate, δ 0 is the reference value of thermal conductivity, σ 0 The content of chlorophyll in the cell fluid of normal healthy plants is 0.5mg / g to 3mg / g, the content of protein is 20mg / g to 30mg / g, the content of chlorogenic acid is 1mg / g to 10mg / g, and the content of jasmonic acid is 10 -6 mg / g to 10 -5 mg / g, chitin content is 10 -8 mg / g, phenylalanine content is 10 -2 mg / g to 5×10 -2 mg / g. In this embodiment, q 1 =3mg / g,q 2 =30mg / g,q 3 =10mg / g,q 4 =10 - 5 mg / g,q 5 =10 -8 mg / g,q 6 =5×10 -2 mg / g, γ 0 =20%,δ 0 =0.4W / (m·K),σ 0 =2.0×10 -7 m² / s. In this embodiment, w1 +w 2 +w 3 +w 4 =1, the first weight, the second weight, the third weight, and the fourth weight are adjusted according to the data of multiple sample plant slices, and w can be set 1 =0.4, w 2 =0.2, w 3 =0.2, w 4 =0.2. As the healthier the plant slices are, the closer the values ​​of light absorption rate, thermal conductivity, and thermal diffusion rate are to the corresponding benchmark values, the disease probability will approach 1, and the probability threshold is set to 0.8.

[0060] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A high-precision plant spectrum detection method based on data fusion, characterized in that: The following steps are involved: Step 1: prepare multiple groups of plant slices, the plant slices enter the first detection channel in turn, the first light generator emits a first laser to the plant slices, the first light receiver collects Raman spectrum data, and the second light receiver collects a first reflection thermal image and a first refraction thermal image; Step 2: extracting corresponding characteristic peaks in the Raman spectrum data according to the disease markers, and calculating a first content set of the disease markers; Step 3: After a preset time, the plant slice enters the second detection channel, and the third optical receiver collects the second reflection thermal image and the second refraction thermal image; Step 4: adjusting the first reflection thermal image, the first refraction thermal image, the second reflection thermal image, and the second refraction thermal image according to the environmental parameters, and then generating the maximum temperature rise parameter, the maximum temperature difference parameter, and the temperature attenuation parameter; Step 5: The plant slice enters the third detection channel, the second light generator emits a second laser to the plant slice, and the fourth light receiver collects infrared spectrum data; Step 6: searching for corresponding characteristic absorption peaks in the infrared spectrum data according to the disease markers, and calculating a second content set of the disease markers; Step 7: predicting the disease probability of the plant slice according to the first content set, the second content set, the maximum temperature rise parameter, the temperature difference parameter and the temperature attenuation parameter of the disease markers; if the disease probability is greater than the probability threshold, marking the plant slice and proceeding to step 8; otherwise, returning to step 1; Step 8: Create a disease classification model based on the Raman spectrum and infrared spectrum of the plant sample, input the Raman spectrum data and infrared spectrum data of the plant slice into the disease classification model, and predict the disease of the plant slice.

2. The high-precision plant spectrum detection method based on data fusion according to claim 1 is characterized in that: The plant slices are citrus fruit slices, the wavelength of the first laser is 785 nm, and the wavelength of the second laser is 10 μm.

3. The high-precision plant spectrum detection method based on data fusion according to claim 1 is characterized in that: In step 2, the disease markers include chlorophyll, protein, chlorogenic acid, jasmonic acid, chitin, and phenylalanine.

4. The high-precision plant spectrum detection method based on data fusion according to claim 1 is characterized in that: In step 4, the environmental parameters include atmospheric temperature T, atmospheric pressure P, relative humidity H, observation distance d, and meteorological visibility V. The atmospheric transmittance η is calculated according to the environmental parameters, and the first reflection thermal image, the first refraction thermal image, the second reflection thermal image, and the second refraction thermal image are adjusted according to the atmospheric transmittance. , a1 is the absorption coefficient, a2 is the scattering coefficient, , , b1 is the first constant, b2 is the second constant, R1 is the specific gas constant of water vapor, T0 is the reference temperature, P0 is the reference pressure, λ is the wavelength of plant thermal radiation, c 1xy =ηc' 1xy , c' 1xy is the pixel value of the first reflection thermal image at the pixel coordinate (x, y), c 1xy is the pixel value of the adjusted first reflected thermal image at the pixel coordinate (x, y).

5. The high-precision plant spectrum detection method based on data fusion according to claim 1 is characterized in that: In step 4, the pixel value of each pixel coordinate in the adjusted first reflection thermal image, the first refraction thermal image, the second reflection thermal image, and the second refraction thermal image is converted into a temperature value, and the maximum temperature rise ΔT1=max{T 1xy -T 2xy }, maximum temperature difference ΔT2=max{|T 1xy -T 2xy |}, T 1xy is the temperature value of the adjusted first reflected thermal image at the pixel coordinate (x, y), T 2xy is the temperature value of the adjusted first refraction thermal image at the pixel coordinate (x, y), the maximum temperature attenuation rate , T 3xy is the temperature value of the adjusted second reflection thermal image at the pixel coordinate (x, y), T 4xy is the temperature value of the adjusted second refraction thermal image at the pixel coordinate (x, y), and t is the preset time length.

6. The high-precision plant spectrum detection method based on data fusion according to claim 5 is characterized in that: In step 7, the first content set = {q 11 ,q 12 ,q 13 ,q 14 ,q 15 ,q 16 }, the second content set = {q 21 ,q 22 ,q 23 ,q 24 ,q 25 ,q 26 }, q 11 is the chlorophyll content obtained from Raman spectroscopy data, q 12 is the protein content obtained from Raman spectroscopy data, q 13 is the chlorogenic acid content obtained from Raman spectroscopy data, q 14 is the jasmonic acid content obtained from Raman spectroscopy data, q 15 is the chitin content obtained from Raman spectroscopy data, q 16 is the phenylalanine content obtained from Raman spectroscopy data, q 21 is the chlorophyll content obtained from infrared spectroscopy data, q 22 is the protein content obtained from infrared spectroscopy data, q 23 is the chlorogenic acid content obtained from infrared spectroscopy data, q 24 is the jasmonic acid content obtained from infrared spectroscopy data, q 25 is the chitin content obtained from infrared spectroscopy data, q 26 Phenylalanine content obtained from infrared spectroscopy data.

7. The high-precision plant spectrum detection method based on data fusion according to claim 6 is characterized in that: In step 7, the light absorption rate is calculated based on the maximum temperature rise, the thermal conductivity is calculated based on the maximum temperature difference, and the thermal diffusion rate is calculated based on the maximum temperature decay. , b3 is the specific heat capacity of the citrus fruit slice, M is the mass of the citrus fruit slice, Δt is the illumination time, P is the incident light power of the first light generator, thermal conductivity , ρ is the heat flux density of the citrus fruit slice, D is the thickness of the citrus fruit slice, and the thermal diffusion rate .

8. The high-precision plant spectrum detection method based on data fusion according to claim 7 is characterized in that: Disease probability , q1 is the benchmark content of chlorophyll, q2 is the benchmark content of protein, q3 is the benchmark content of chlorogenic acid, q4 is the benchmark content of jasmonic acid, q5 is the benchmark content of chitin, q6 is the benchmark content of phenylalanine, w1 is the first weight, w2 is the second weight, w3 is the third weight, w4 is the fourth weight, γ0 is the benchmark value of light absorption rate, δ0 is the benchmark value of thermal conductivity, and σ0 is the benchmark value of thermal diffusion rate.

9. The high-precision plant spectrum detection method based on data fusion according to claim 1, characterized in that: In step 8, Raman spectra and infrared spectra of multiple groups of plant samples are collected, and the disease type and disease area are marked on each Raman spectrum and infrared spectrum. The marked Raman spectrum and infrared spectrum are divided into a training set and a validation set. The disease classification model is trained using a supervised deep learning method using the training set, and the training error is calculated according to the loss function. Then, at least one parameter of the disease classification model is adjusted according to the training error. Finally, the validation set is used to verify and output the disease classification model.

Images