Citrus huanglongbing detection device and method based on near infrared characteristic spectrum

Abstract

The invention provides a Citrus Huanglongbing detection device and method based on near-infrared characteristic spectrum.The method comprises the steps that citrus leaves are processed through an optical acquisition module and a signal processing module, and digital signals are obtained; performing standardization processing on the digital signal to obtain a sample characteristic spectrum, and inputting the sample characteristic spectrum into a data enhancement model to generate a virtual characteristic spectrum so as to expand a sample set; inputting the sample characteristic spectrum and the virtual characteristic spectrum into a detection model, and generating a judgment result of the Candidatus Liberobacter asiaticum; according to the invention, whether the citrus suffers from the citrus huanglongbing or not can be conveniently detected on site in real time.

Description

Technical Field

[0001] This invention relates to the field of agricultural testing technology, and in particular to a device and method for detecting citrus Huanglongbing (HLB) based on near-infrared characteristic spectroscopy. Background Technology

[0002] Citrus, as one of the world's most widely cultivated economic crops, is susceptible to hundreds of diseases and pests during its growth period. Among these, HLB (Huanglongbing) is the most serious, often referred to as the "cancer" of citrus, severely hindering the industry's development. This devastating disease is caused by Gram-negative bacteria infecting the phloem of the plant and is characterized by its extremely rapid spread. Currently, there is no effective cure; timely uprooting and burning of infected plants is the most effective control method. Therefore, early detection of HLB is particularly urgent. Current detection methods fall into two main categories: field diagnosis and laboratory biochemical analysis. The former mainly diagnoses based on the symptoms exhibited by citrus plants, which is simple and easy to perform, but requires a high level of experience, is subjective, and has low efficiency. The latter includes physicochemical index detection methods and PCR detection methods, which are complex, cumbersome, and time-consuming, requiring a high level of professional knowledge from the testers, and are not easily applicable to large-scale citrus orchard production. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide a method for detecting Huanglongbing (HLB) in citrus based on near-infrared characteristic spectroscopy, thereby resolving the technical problems mentioned in the background section.

[0004] To achieve the above objectives, the present invention is implemented through the following technical solution: A method for detecting Huanglongbing (HLB) in citrus based on near-infrared characteristic spectroscopy includes the following steps: The citrus leaves are processed by an optical acquisition module and a signal processing module to obtain digital signals; The digital signal is standardized to obtain the sample feature spectrum, and the sample feature spectrum is input into the data augmentation model to generate a virtual feature spectrum to expand the sample set; The sample characteristic spectrum and the virtual characteristic spectrum are input into the detection model to generate the discrimination result of citrus Huanglongbing.

[0005] According to one aspect of the above technical solution, the optical acquisition module includes a light source, a beam splitter, and a photoelectric sensor; the signal processing module includes a preamplifier circuit, a filter circuit, and an A / D converter; the light source is used to illuminate citrus leaves; the beam splitter is used to generate light in a sensitive wavelength range and filter out irrelevant wavelength range light; the photoelectric sensor is used to collect near-infrared reflected light from the citrus leaves; the preamplifier circuit is used to convert the photocurrent signal into a voltage signal; the filter circuit is used to suppress signals outside a specific frequency range; and the A / D converter is used to convert the voltage signal into a digital signal.

[0006] According to one aspect of the above technical solution, the step of standardizing the digital signal to obtain the sample feature spectrum and inputting the sample feature spectrum into a data augmentation model to generate a virtual feature spectrum to expand the sample set includes: The digital signal is standardized to obtain the sample characteristic spectrum; An adversarial generative network is constructed, and the sample feature spectrum is input into the discriminator of the adversarial generative network to train the discriminator; The random noise signal and the original digital signal are input into the generator of the adversarial generative network to generate a virtual feature spectrum; The virtual feature spectrum is input into the discriminator, and the generator continues to generate virtual feature spectra to train the discriminator and the generator alternately. When the adversarial generative network reaches Nash equilibrium, the generator uses the generated virtual feature spectra as data augmentation samples to expand the sample set.

[0007] According to one aspect of the above technical solution, the loss function of the data augmentation model during training is as follows: ; Where D represents the discriminator, G represents the generator, minG represents minimizing the generator, maxD represents maximizing the discriminator, V(D,G) represents the value function, E represents the expected value, y represents the label of a batch of sample characteristic spectra, and x represents the characteristic spectrum of a single sample in a batch of sample characteristic spectra. This indicates that a sample characteristic spectrum x is sampled from the characteristic spectra of the batch of samples labeled y, and D(x,y) represents the score given by the discriminator to the sample characteristic spectrum. Let z represent the virtual feature spectrum randomly sampled from the noise distribution, and G(z,y) represent the virtual feature spectrum generated by the generator.

[0008] According to one aspect of the above technical solution, the structure of the adversarial generative network sequentially includes an implantation layer, a first tiling layer, a product layer, a first fully connected layer, a first convolutional layer, a linear rectified activation function layer, a first batch normalization layer, an upper sampling layer, a second convolutional layer, a hyperbolic tangent function layer, a third convolutional layer, a leaky linear rectified activation layer, a random deactivation layer, a second batch normalization layer, a second tiling layer, a second fully connected layer, and an output layer.

[0009] According to one aspect of the above technical solution, the specific steps of inputting the sample characteristic spectrum and the virtual characteristic spectrum into the detection model to generate the discrimination result of citrus Huanglongbing include: A detection model is constructed based on a width learning network. The sample feature spectrum and the virtual feature spectrum are used as sample sets and input into the feature layer of the width learning network to obtain the output matrix Z of each feature mapping. i ; ; Where i represents the sequence number of the feature map group, ξ represents the linear activation function, and W i f B represents a randomly generated weight matrix. i f X represents the randomly generated bias vector, FM represents the total number of feature maps, and X represents the total number of feature maps. train The sample characteristic spectrum and the virtual characteristic spectrum are represented in the sample set; The output matrices of each set of feature maps are concatenated to obtain the feature layer matrix Z; ; Calculate the Euclidean distance d between any two samples in the feature layer matrix. ab To obtain the elements in the kernel matrix. ; ; ; in, Represents the kernel matrix of the first... a row and number b The elements of the column, Z a and Z b They represent the first, second, and third features in the feature layer matrix Z, respectively. a row and number b The vector of the row, corresponding to the first row in the sample set. a The and the first b The feature vector of a sample, where N represents the total number of samples in the sample set, and σ represents the kernel width parameter; The kernel matrix is ​​obtained by concatenating all elements of the kernel matrix. ; The feature layer matrix and the kernel matrix are concatenated to obtain the first hidden layer H; ; Provide a reference category label Y, calculate the output weight matrix W, and complete the training phase; ; Where λ represents the regularization parameter and I represents the identity matrix; Acquire real-time spectral data X of the sample to be tested. test ; Using the random weights and biases saved during the training phase, we obtain the individual feature map Z. test，i By concatenating all the individual feature maps, the total feature map Z is obtained. test ; ; ; Calculate the Euclidean distance d between the test sample and each sample in the training phase. test，k , obtain kernel similarity Then calculate the kernel vector. ; ; ; ; The second hidden layer H is obtained based on the total feature map and the kernel vector. test ; ; Based on the output weight matrix of the second hidden layer and the training phase, the predicted value is calculated. ; ; Based on the predicted values, determine whether the citrus fruit is infected with Huanglongbing (HLB); ; Where, p low P represents the lower limit threshold for disease incidence. high This indicates the upper limit threshold for the incidence of the disease.

[0010] According to one aspect of the above technical solution, the light source is a near-infrared LED, the beam splitter is a rotating filter, the photoelectric sensor is a miniature InGaAs detector, the preamplifier circuit is a transimpedance amplifier, and the filter circuit is a bandpass filter.

[0011] This invention also provides a citrus Huanglongbing (HLB) detection device based on near-infrared characteristic spectra, comprising a housing, a sample chamber disposed on the housing, an optical acquisition module disposed in the sample chamber, a signal processing module, and a detection module. The optical acquisition module includes a light source, a spectrometer, and a photoelectric sensor disposed in the sample chamber. The signal processing module includes a preamplifier circuit, a filter circuit, and an A / D converter. The detection module includes a data enhancement model and a detection model. The spectrometer is used to generate light of a sensitive wavelength band from the light source and filter out irrelevant wavelength band light. The photoelectric sensor is used to collect near-infrared reflected light from citrus leaves. The preamplifier circuit is used to convert the photocurrent signal into a voltage signal. The filter circuit is used to suppress signals outside a specific frequency range. The A / D converter is used to convert the voltage signal into a digital signal. The data augmentation model serves the following purposes: The digital signal is standardized to obtain the sample feature spectrum, and the sample feature spectrum is input into the data augmentation model to generate a virtual feature spectrum to expand the sample set; The detection model serves the following purpose: The sample characteristic spectrum and the virtual characteristic spectrum are input into the detection model to generate the discrimination result of citrus Huanglongbing.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: By designing an outer casing with a sample chamber on it, citrus leaves are placed inside for testing. The system is simple to operate and portable, meeting the needs of large-scale, real-time on-site screening in citrus orchards. An optical acquisition module is installed in the sample chamber. Before use, citrus leaves are placed inside, and light from an LED light source is filtered through a rotating filter to produce light of a specific wavelength that illuminates the leaves. The reflected light signal is collected by a photoelectric sensor and converted into an electrical signal. A preamplifier circuit converts the weak photocurrent signal into a voltage signal, and a filter circuit suppresses or attenuates signals outside a specific frequency range. An A / D converter then converts the voltage signal into a digital signal. Finally, a generative adversarial network (GAN) is constructed, allowing the discriminator and generator to interact. In the game theory approach, when Nash equilibrium is reached—that is, when both the discriminator and the generator make optimal choices and are in a stable state—the generator will produce virtual feature spectra that are highly similar to the real feature spectra and have accurate labels. This expands the spectrum and labels, providing a large number of high-quality samples for subsequent model construction. This effectively solves the problem of insufficient training of deep learning models and difficulty in improving prediction accuracy and stability due to insufficient sample size. Then, by constructing a training set from the virtual feature spectra and the sample feature spectra, and inputting it into a width learning network, the health status of citrus fruits is classified. This device can then be used to detect citrus leaves at any time to determine whether citrus fruits have Huanglongbing (HLB). The model parameters will also be updated periodically to optimize the model. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the citrus Huanglongbing detection device based on near-infrared characteristic spectrum in the first embodiment of the present invention from a first-view perspective; Figure 2 A schematic diagram of the citrus Huanglongbing detection device based on near-infrared characteristic spectrum in the first embodiment of the present invention from a second perspective; Figure 3 This is a structural block diagram of the data augmentation model in the first embodiment of the present invention; Figure 4 The flowchart of the citrus Huanglongbing detection method based on near-infrared characteristic spectrum in the second embodiment of the invention; Explanation of key component symbols:

[0014] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0015] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0016] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0018] Please see Figures 1 to 2The image shows a citrus Huanglongbing (HLB) detection device based on near-infrared characteristic spectroscopy according to a first embodiment of the present invention. It includes a housing 1, a sample chamber 6 disposed on the housing, an optical acquisition module disposed in the sample chamber, a signal processing module, and a detection module. The optical acquisition module includes a light source, a beam splitter, and a photoelectric sensor disposed in the sample chamber. The signal processing module includes a preamplifier circuit, a filter circuit, and an A / D converter. The detection module includes a data enhancement model and a detection model. The beam splitter is used to generate light in a sensitive wavelength band from the light source and filter out irrelevant wavelength band light. The photoelectric sensor is used to collect near-infrared reflected light from citrus leaves. The preamplifier circuit is used to convert the photocurrent signal into a voltage signal. The filter circuit is used to suppress signals outside a specific frequency range. The A / D converter is used to convert the voltage signal into a digital signal. The data augmentation model serves the following purposes: The digital signal is standardized to obtain the sample feature spectrum, and the sample feature spectrum is input into the data augmentation model to generate a virtual feature spectrum to expand the sample set; The detection model serves the following purpose: The sample characteristic spectrum and the virtual characteristic spectrum are input into the detection model to generate the discrimination result of citrus Huanglongbing.

[0019] The steps include: standardizing the digital signal to obtain the sample feature spectrum, and inputting the sample feature spectrum into the data augmentation model to generate a virtual feature spectrum to expand the sample set. The digital signal is standardized to obtain the sample characteristic spectrum; An adversarial generative network is constructed, and the sample feature spectrum is input into the discriminator of the adversarial generative network to train the discriminator; The random noise signal and the original digital signal are input into the generator of the adversarial generative network to generate a virtual feature spectrum; The virtual feature spectrum is input into the discriminator, and the generator continues to generate virtual feature spectra to train the discriminator and the generator alternately. When the adversarial generative network reaches Nash equilibrium, the generator uses the generated virtual feature spectra as data augmentation samples to expand the sample set.

[0020] Understandably, this invention, by setting up an outer shell and a sample chamber on it, allows citrus leaves to be placed in the sample chamber for testing. It is simple to operate and easy to carry, meeting the needs of large-scale, real-time on-site screening in citrus orchards. An optical acquisition module is installed in the sample chamber. Before use, citrus leaves are placed in the sample chamber. Light generated by an LED light source passes through a rotating filter to produce light of a specific wavelength that illuminates the citrus leaves. The reflected light signal is collected by a photoelectric sensor and converted into an electrical signal. A preamplifier circuit converts the weak photocurrent signal into a voltage signal. A filter circuit suppresses or attenuates signals outside a specific frequency range. An A / D converter converts the voltage signal into a digital signal. Then, by constructing a generative adversarial network, the discriminator and... The generators engage in a game-like process. When a Nash equilibrium is reached—meaning both the discriminator and the generator have made optimal choices and are in a stable state—the generator produces virtual feature spectra that are highly similar to the real feature spectra and have accurate labels. This expands the spectrum and labels, providing a large number of high-quality samples for subsequent model construction. This effectively solves the problem of insufficient training of deep learning models and difficulty in improving prediction accuracy and stability due to insufficient sample size. Then, by constructing a training set from the virtual feature spectra and the sample feature spectra, the data is input into a width learning network to classify the health status of citrus fruits. This device can then be used to detect citrus leaves at any time to determine whether citrus fruits have Huanglongbing (HLB). The model parameters will be updated periodically to optimize the model.

[0021] Specifically, the loss function of the data augmentation model during training is as follows: ; Where D represents the discriminator, G represents the generator, minG represents minimizing the generator, maxD represents maximizing the discriminator, V(D,G) represents the value function, E represents the expected value, y represents the label of a batch of sample characteristic spectra, and x represents the characteristic spectrum of a single sample in a batch of sample characteristic spectra. This indicates that a sample characteristic spectrum x is sampled from the characteristic spectra of the batch of samples labeled y, and D(x,y) represents the score given by the discriminator to the sample characteristic spectrum. Let z represent the virtual feature spectrum randomly sampled from the noise distribution, and G(z,y) represent the virtual feature spectrum generated by the generator.

[0022] Understandably, maximizing the discriminator, that is, the discriminator D distinguishing real data from generated fake data as much as possible, makes D(x,y) approach 1 (identifying real data) and D(G(z,y),y) approach 0 (identifying fake data). When the discriminator D's capability is optimal, and To reach the maximum value, minimize the generator, that is, make the generator G "deceive" the discriminator as much as possible to misclassify the generated data as real data, i.e., make D(G(z,y),y) approach 1. This process is equivalent to minimizing The value of the value function reflects the discriminant performance level of the discriminator D, where the discriminator D aims to maximize the value of this function, while the generator G aims to reduce the value of this function.

[0023] Furthermore, such as Figure 3 As shown, the structure of the adversarial generative network includes, in sequence, an implantation layer, a first tiling layer, a product layer, a first fully connected layer, a first convolutional layer, a linear rectified activation function layer, a first batch normalization layer, an upper sampling layer, a second convolutional layer, a hyperbolic tangent function layer, a third convolutional layer, a leaky linear rectified activation function layer, a random deactivation layer, a second batch normalization layer, a second tiling layer, a second fully connected layer, and an output layer.

[0024] Understandably, the first half of a generative adversarial network (GAN) is a generator, and the second half is a discriminator. The generator has two inputs: a normally distributed random noise signal and the original digital signal of the sampled label data. The sampled labels are fitted based on real label data. After passing through an implantation layer and a first tiling layer, they are multiplied with the noise data to obtain the network input. This input is then processed through a first fully connected layer and multiple convolutional layers to output a virtual spectrum with realistic features as the generated spectrum. A linear rectified activation function layer is connected after the first convolutional layer to enhance the network's nonlinear processing capability. The second convolutional layer is connected to a hyperbolic tangent function layer to ensure the network output data is reasonable. A first batch normalization layer and an upper sampling layer are inserted between the two convolutional layers to ensure stable feature transformation and spatial expansion of the generator network. This hierarchical transformation effectively strengthens the generator network's ability to process and generate disordered data. The discriminator's input consists of the virtual feature spectrum generated by the generator and the original sample feature spectrum. Its third convolutional layer is connected to a leaky linear rectified activation layer, which enhances the network's nonlinear processing capability. A random deactivation layer is connected to a second batch normalization layer to ensure the discriminator's stable and continuous learning ability and improve the model's generalization ability. When the generator and discriminator reach "Nash equilibrium," that is, when both the generator and discriminator make optimal choices and reach a stable state, the feature spectra generated by the generator are considered sufficient as virtual samples to be used as input for subsequent detection model training.

[0025] Furthermore, the specific steps of inputting the sample characteristic spectrum and the virtual characteristic spectrum into the detection model to generate the discrimination result of citrus Huanglongbing include: A detection model is constructed based on a width learning network. The sample feature spectrum and the virtual feature spectrum are used as sample sets and input into the feature layer of the width learning network to obtain the output matrix Z of each feature mapping.i ; ; Where i represents the sequence number of the feature map group, ξ represents the linear activation function, and W i f B represents a randomly generated weight matrix. i f X represents the randomly generated bias vector, FM represents the total number of feature maps, and X represents the total number of feature maps. train The sample characteristic spectrum and the virtual characteristic spectrum are represented in the sample set; The output matrices of each set of feature maps are concatenated to obtain the feature layer matrix Z; ; Calculate the Euclidean distance d between any two samples in the feature layer matrix. ab To obtain the elements in the kernel matrix. ; ; ; in, Represents the kernel matrix of the first... a row and number b The elements of the column, Z a and Z b They represent the first, second, and third features in the feature layer matrix Z, respectively. a row and number b The vector of the row, corresponding to the first row in the sample set. a The and the first b The feature vector of a sample, where N represents the total number of samples in the sample set, and σ represents the kernel width parameter; The kernel matrix is ​​obtained by concatenating all elements of the kernel matrix. ; The feature layer matrix and the kernel matrix are concatenated to obtain the first hidden layer H; ; Provide a reference category label Y, calculate the output weight matrix W, and complete the training phase; ; Where λ represents the regularization parameter and I represents the identity matrix; Acquire real-time spectral data X of the sample to be tested. test ; Using the random weights and biases saved during the training phase, we obtain the individual feature map Z. test，i By concatenating all the individual feature maps, the total feature map Z is obtained. test ; ; ; Calculate the Euclidean distance d between the test sample and each sample in the training phase. test，k , obtain kernel similarity Then calculate the kernel vector. ; ; ; ; The second hidden layer H is obtained based on the total feature map and the kernel vector. test ; ; Based on the output weight matrix of the second hidden layer and the training phase, the predicted value is calculated. ; ; Based on the predicted values, determine whether the citrus fruit is infected with Huanglongbing (HLB); ; Where, p low P represents the lower limit threshold for disease incidence. high This indicates the upper limit threshold for the incidence of the disease.

[0026] Understandably, the detection model is based on a wide-learning network. However, existing discriminative models involve a double random mapping from the input layer to the feature layer and then to the enhancement layer, which leads to multiple training sessions on the same training data, resulting in different models and large performance fluctuations. This application uses a kernel matrix instead of random enhancement nodes. Given the feature layer matrix Z and the kernel width parameter σ, the kernel matrix is ​​deterministic, ensuring that the training results are completely consistent and repeatable. Furthermore, the feature layer matrix Z is projected onto an infinite-dimensional feature space through the kernel, making the originally non-linearly separable problem linearly separable. Due to the regularization properties of the kernel matrix, this model has a low risk of overfitting and is easy to deploy with small sample learning, thereby improving the accuracy and portability of citrus Huanglongbing detection.

[0027] Furthermore, the light source is a ring array of highly stable near-infrared LEDs, covering the characteristic absorption band of citrus leaves; the beam splitter is a rotating filter that generates light in the sensitive band and filters out irrelevant or correlated low-band light; the photoelectric sensor is a miniature InGaAs detector that supports multi-channel data and is used to collect the near-infrared reflected light of the target object. After collection, the data is sent to the signal processing module.

[0028] Furthermore, the preamplifier circuit is a low-noise transimpedance amplifier used to convert weak photocurrent signals into voltage signals, minimizing noise introduction during the conversion process; the filter circuit is a bandpass filter to suppress or attenuate signals outside a specific frequency range; the A / D converter uses a 16-bit high-precision ADC, whose main function is to convert analog signals into 16-bit digital signals and perform standardization processing in conjunction with whiteboard calibration values ​​for processing by the main control chip. This module receives data from the optical acquisition module and filters out interference from non-target bands.

[0029] Furthermore, the citrus Huanglongbing detection device based on near-infrared characteristic spectrum also includes a main control chip MCU and an ESP32-S3 communication module mounted on the outer casing. The main control chip MCU is electrically connected to the optical acquisition module, signal processing module, and detection module. The ESP32-S3 communication module is electrically connected to the detection module and is used to upload the detection results.

[0030] Understandably, the main control chip is a low-power MCU (STM32H7 series, supporting floating-point operations). The detection model is stored in the main control chip MCU. The main control chip MCU can control the light source to emit light with a wavelength range of 400-1050nm, and simultaneously control the rotation of the rotating filter to select light of sensitive wavelength bands (493, 515, 630, 700, 716, 739, 771, and 958) to illuminate the citrus leaves in the sample chamber, showing significant differences in the characteristic spectra of normal and Huanglongbing leaves. Among them, chlorophyll degradation leads to a decrease in reflectance at 493 and 515 nm (blue and green light bands) and an increase in reflectance at 700 nm (red light band); cell structure damage leads to a decrease in reflectance at 716, 739, and 771 nm (near-infrared light bands); and obstructed water transport leads to an increase in reflectance at 958 nm. The ESP32-S3 communication module connects to the network via Wi-Fi, transmitting virtual samples and diagnostic results sent by the detection model in the main control chip's MCU to the cloud server. The cloud server further processes and stores the data, trains the model in the cloud to improve discrimination accuracy, and then feeds the updated model back to the detection device through the communication module. Additionally, it can connect to a mobile app via Bluetooth BLE 5.0, allowing users to conveniently view the detection results on their mobile devices.

[0031] Furthermore, the citrus Huanglongbing detection device based on near-infrared characteristic spectrum also includes multiple control buttons, a MicroSD card slot 3, a USB interface 4, a power button 5, and a display screen 7 on the outer casing. The MicroSD card slot is used to install a MicroSD memory card, the USB interface is used to connect to the ESP32-S3 communication module, and the display screen is electrically connected to the detection module.

[0032] Understandably, multiple control buttons include: Sample chamber knob 8: locks the sample chamber; Spectral magnification / reduction knob 9: left-hand rotation magnifies, right-hand rotation reduces, facilitating observation of spectral characteristic peaks and other details; Detection button 10: long press for calibration, short press to start detection after calibration, mainly calculating the energy of each band and saving it as a comparison reference value to obtain the sample's characteristic spectrum; Confirm button 11: confirms the save and upload buttons on the interactive interface; Save button 12: saves the original characteristic spectrum, selectable to save to the built-in FLASH memory or MicroSD card; Upload button 13: uploads data and updates the model, selectable to upload to cloud service or mobile APP; Movement button 14: selects different sample characteristic spectra for display. The cloud server receives the uploaded data, updates the database, and uses data augmentation technology to retrain and optimize the model, selecting the best model and covering it.

[0033] Preferably, the outer shell adopts a waterproof and dustproof box-type structure, which is convenient for storage, handling and carrying; the main body of the device is connected to an ergonomic grip handle, which can also prevent the built-in battery from becoming loose; the sample chamber on the front of the device adopts a top-open pull-out cuboid structure that can accommodate multiple citrus leaves at the same time, and a knob is provided on the right side of the sample chamber to lock the sample chamber, so as to keep the sample from being affected by the external environment during the operation of the device.

[0034] The operating procedure for this device is as follows: Press the power button to initiate a self-test, verifying the proper functioning of the optical acquisition module, signal processing module, and communication module. If any abnormalities are detected, an error code will be displayed on the screen, and the detection model will be loaded simultaneously. After the self-test is complete, press and hold the detection button. The system will trigger LEDs to illuminate sequentially, acquiring the reflectance spectrum of the white board, calculating the energy of each band, and storing it as a reference value. The screen will then display "Calibration complete, detection can begin." Select the citrus tree to be tested, pick a representative leaf, place it in the sample chamber, press the detection button, acquire the reflected light from the sample, and calculate the reflectance, comparing it with the reference value. This will calculate the current leaf spectrum and display it on the screen. Simultaneously, compare it with the loaded detection model to analyze and calculate whether the leaf suffers from Huanglongbing (healthy / suspected / confirmed). The screen will display the result, and the save button can be pressed to save the information.

[0035] If the detection model is not loaded, the original and preprocessed feature spectra, along with a timestamped JSON file, are stored in the built-in storage to form the original dataset. After synchronizing the obtained data to the cloud server via the upload button, a large number of training samples are obtained through data augmentation models based on this original dataset. Deep learning training is then performed to optimize and obtain a better detection model, which is imported into the main control chip (MCU) to overwrite the previous detection model, thereby enabling on-site detection to diagnose the health status of the tested leaves.

[0036] In summary, the citrus Huanglongbing (HLB) detection device based on near-infrared characteristic spectroscopy in the above embodiments of the present invention integrates modern spectral analysis methods and possesses functions such as data acquisition, data transmission, and display. It optimizes eight sensitive wavelengths through rotating filters and coordinates with the cloud, combining data augmentation technology to train and optimize the model, aiming to obtain the best model and achieve rapid, non-destructive field diagnosis of HLB. Furthermore, the device and method are simple, support real-time data display, data synchronization with the cloud, and dynamic model updates, significantly improving orchard management efficiency and profitability, and providing important practical guidance for citrus production.

[0037] Please refer to Figure 4 The image shows a method for detecting citrus Huanglongbing based on near-infrared characteristic spectroscopy in a second embodiment of the present invention, comprising the following steps: S10 processes the citrus leaves through an optical acquisition module and a signal processing module to obtain digital signals; S20, the digital signal is standardized to obtain the sample feature spectrum, and the sample feature spectrum is input into the data augmentation model to generate a virtual feature spectrum to expand the sample set; S30, input the sample characteristic spectrum and the virtual characteristic spectrum into the detection model to generate the discrimination result of citrus Huanglongbing.

[0038] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0039] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for detecting citrus Huanglongbing (HLB) based on near-infrared characteristic spectroscopy, characterized in that, Includes the following steps: The citrus leaves are processed by an optical acquisition module and a signal processing module to obtain digital signals; The digital signal is standardized to obtain the sample feature spectrum, and the sample feature spectrum is input into the data augmentation model to generate a virtual feature spectrum to expand the sample set; The sample characteristic spectrum and the virtual characteristic spectrum are input into the detection model to generate the discrimination result of citrus Huanglongbing.

2. The method for detecting citrus Huanglongbing based on near-infrared characteristic spectroscopy according to claim 1, characterized in that, The optical acquisition module includes a light source, a beam splitter, and a photoelectric sensor. The signal processing module includes a preamplifier circuit, a filter circuit, and an A / D converter. The light source is used to illuminate citrus leaves. The beam splitter is used to generate light in a sensitive wavelength range and filter out irrelevant wavelength range light. The photoelectric sensor is used to collect near-infrared reflected light from the citrus leaves. The preamplifier circuit is used to convert the photocurrent signal into a voltage signal. The filter circuit is used to suppress signals outside a specific frequency range. The A / D converter is used to convert the voltage signal into a digital signal.

3. The method for detecting citrus Huanglongbing based on near-infrared characteristic spectroscopy according to claim 2, characterized in that, The steps include: standardizing the digital signal to obtain the sample feature spectrum, and inputting the sample feature spectrum into the data augmentation model to generate a virtual feature spectrum to expand the sample set. The digital signal is standardized to obtain the sample characteristic spectrum; An adversarial generative network is constructed, and the sample feature spectrum is input into the discriminator of the adversarial generative network to train the discriminator; The random noise signal and the original digital signal are input into the generator of the adversarial generative network to generate a virtual feature spectrum; The virtual feature spectrum is input into the discriminator, and the generator continues to generate virtual feature spectra to train the discriminator and the generator alternately. When the adversarial generative network reaches Nash equilibrium, the generator uses the generated virtual feature spectra as data augmentation samples to expand the sample set.

4. The method for detecting citrus Huanglongbing based on near-infrared characteristic spectroscopy according to claim 3, characterized in that, The loss function for training the data augmentation model is as follows: ； Where D represents the discriminator, G represents the generator, minG represents minimizing the generator, maxD represents maximizing the discriminator, V(D,G) represents the value function, E represents the expected value, y represents the label of a batch of sample characteristic spectra, and x represents the characteristic spectrum of a single sample in a batch of sample characteristic spectra. This indicates that a sample characteristic spectrum x is sampled from the characteristic spectra of the batch of samples labeled y, and D(x,y) represents the score given by the discriminator to the sample characteristic spectrum. Let z represent the virtual feature spectrum randomly sampled from the noise distribution, and G(z,y) represent the virtual feature spectrum generated by the generator.

5. The method for detecting citrus Huanglongbing based on near-infrared characteristic spectroscopy according to claim 4, characterized in that, The structure of the adversarial generative network includes, in sequence, an implantation layer, a first tiling layer, a product layer, a first fully connected layer, a first convolutional layer, a linear rectified activation function layer, a first batch normalization layer, an upper sampling layer, a second convolutional layer, a hyperbolic tangent function layer, a third convolutional layer, a leaky linear rectified activation function layer, a random deactivation layer, a second batch normalization layer, a second tiling layer, a second fully connected layer, and an output layer.

6. The method for detecting citrus Huanglongbing based on near-infrared characteristic spectroscopy according to claim 1, characterized in that, The specific steps of inputting the sample characteristic spectrum and the virtual characteristic spectrum into the detection model to generate the discrimination result of citrus Huanglongbing include: A detection model is constructed based on a width learning network. The sample feature spectrum and the virtual feature spectrum are used as sample sets and input into the feature layer of the width learning network to obtain the output matrix Z of each feature mapping. i ; ； Where i represents the sequence number of the feature map group, ξ represents the linear activation function, and W i f B represents a randomly generated weight matrix. i f X represents the randomly generated bias vector, FM represents the total number of feature maps, and X represents the total number of feature maps. train The sample characteristic spectrum and the virtual characteristic spectrum are represented in the sample set; The output matrices of each set of feature maps are concatenated to obtain the feature layer matrix Z; ； Calculate the Euclidean distance d between any two samples in the feature layer matrix. ab To obtain the elements in the kernel matrix. ; ； ； in, Represents the kernel matrix of the first... a row and number b The elements of the column, Z a and Z b They represent the first, second, and third features in the feature layer matrix Z, respectively. a row and number b The vector of the row, corresponding to the first row in the sample set. a The and the first b The feature vector of a sample, where N represents the total number of samples in the sample set, and σ represents the kernel width parameter; The kernel matrix is ​​obtained by concatenating all elements of the kernel matrix. ; The feature layer matrix and the kernel matrix are concatenated to obtain the first hidden layer H; ； Provide a reference category label Y, calculate the output weight matrix W, and complete the training phase; ； Where λ represents the regularization parameter and I represents the identity matrix; Acquire real-time spectral data X of the sample to be tested. test ; Using the random weights and biases saved during the training phase, we obtain the individual feature map Z. test，i By concatenating all the individual feature maps, the total feature map Z is obtained. test ; ； ； Calculate the Euclidean distance d between the test sample and each sample in the training phase. test，k , obtain kernel similarity Then calculate the kernel vector. ; ； ； ； The second hidden layer H is obtained based on the total feature map and the kernel vector. test ; ； Based on the output weight matrix of the second hidden layer and the training phase, the predicted value is calculated. ; ； Based on the predicted values, determine whether the citrus fruit is infected with Huanglongbing (HLB); ； Where, p low P represents the lower limit threshold for the disease. high This indicates the upper limit threshold for the incidence of the disease.

7. The method for detecting citrus Huanglongbing based on near-infrared characteristic spectroscopy according to claim 2, characterized in that, The light source is a near-infrared LED, the beam splitter is a rotating filter, the photoelectric sensor is a miniature InGaAs detector, the preamplifier circuit is a transimpedance amplifier, and the filter circuit is a bandpass filter.

8. A citrus Huanglongbing (HLB) detection device based on near-infrared characteristic spectroscopy, characterized in that, The device includes a housing, a sample chamber mounted on the housing, an optical acquisition module disposed within the sample chamber, a signal processing module, and a detection module. The optical acquisition module includes a light source, a beam splitter, and a photoelectric sensor disposed within the sample chamber. The signal processing module includes a preamplifier circuit, a filter circuit, and an A / D converter. The detection module includes a data enhancement model and a detection model. The beam splitter is used to generate light of a sensitive wavelength band from the light source and filter out irrelevant wavelength band light. The photoelectric sensor is used to collect near-infrared reflected light from citrus leaves. The preamplifier circuit is used to convert the photocurrent signal into a voltage signal. The filter circuit is used to suppress signals outside a specific frequency range. The A / D converter is used to convert the voltage signal into a digital signal. The data augmentation model serves the following purposes: The digital signal is standardized to obtain the sample feature spectrum, and the sample feature spectrum is input into the data augmentation model to generate a virtual feature spectrum to expand the sample set; The detection model serves the following purpose: The sample characteristic spectrum and the virtual characteristic spectrum are input into the detection model to generate the discrimination result of citrus Huanglongbing.

Images