A tobacco virus classification model construction method based on tobacco hyperspectral

Abstract

The application provides a tobacco virus classification model construction method based on tobacco leaf hyperspectrum, and belongs to the technical field of tobacco virus classification models. Fractional differential spectrum enhancement processing is adopted, persistent homology eigenvalues and Betti number sequence values are calculated through spectrum topological manifold embedding of enhanced spectrum data, dimension reduction is performed through kernel principal component analysis to obtain topological embedding feature vectors, fusion feature vectors are output through a multimodal attention fusion module in combination with thermal infrared temperature field data and chlorophyll fluorescence kinetic curve data, data enhancement is performed on minority class samples by using a spectral generative adversarial network, the fusion feature vectors are input into a virus classification and recognition model adopting a focal loss function and a dynamic network adjustment mechanism for training, virus type classification and disease severity grade recognition of tobacco leaves to be tested are realized, and the problem that different types of tobacco viruses cannot be accurately recognized and distinguished in the early stage of disease is solved.

Description

Technical Field

[0001] This invention belongs to the technical field of tobacco virus classification models, and more specifically, relates to a method for constructing a tobacco virus classification model based on tobacco leaf hyperspectral imaging. Background Technology

[0002] Tobacco viral diseases are significant factors affecting tobacco leaf yield and quality. Traditional detection methods primarily rely on hyperspectral reflectance analysis to determine disease types by measuring the reflectance characteristics of leaves in the visible to near-infrared bands. However, existing spectral analysis methods only focus on point-by-point numerical changes in the spectral curve, extracting disease features by comparing reflectance differences in specific bands or calculating first-order derivative spectra. This feature extraction method, based on local numerical changes, struggles to capture the global geometric structure information of the spectral curve and exhibits poor robustness to changes in ambient light and measurement noise. In the early stages of disease, the subtle changes in chlorophyll content and minor cell structural damage caused by viral infection are often masked by instrument noise on the spectral curve. Traditional feature extraction methods cannot effectively amplify these subtle differences, and the spectral features of different virus types lack sufficient differentiation in the early stages, leading to a high misclassification rate. In other words, existing technologies suffer from the technical challenge of accurately identifying and distinguishing different types of tobacco viruses in the early stages of disease. Summary of the Invention

[0003] In view of this, the present invention provides a method for constructing a tobacco virus classification model based on tobacco leaf hyperspectral imaging, which can solve the technical problem in the prior art that it is difficult to accurately identify and distinguish different types of tobacco viruses in the early stages of disease.

[0004] This invention is implemented as follows: It provides a method for constructing a tobacco virus classification model based on tobacco leaf hyperspectral data. This includes collecting hyperspectral reflectance data of tobacco leaves in the wavelength range of 380nm to 1000nm, performing Savitzky-Golay filtering to denoise the data, obtaining denoised hyperspectral reflectance data, performing fractional-order differential spectral enhancement on the denoised hyperspectral reflectance data to obtain fractional-order differential enhanced spectral data, performing spectral topological manifold embedding on the enhanced spectral data to calculate persistent homology eigenvalues ​​and Betti number sequence values, and obtaining a topological embedding feature vector through dimensionality reduction using kernel principal component analysis. It also involves collecting thermal infrared temperature field data and chlorophyll fluorescence dynamics curve data corresponding to tobacco leaves, inputting the topological embedding feature vector, thermal infrared temperature field data, and chlorophyll fluorescence dynamics curve data into a multimodal attention fusion module to obtain a fused feature vector, counting the number of samples in each class, and generating a synthetic fused feature vector for minority class samples using a spectral generative adversarial network, and inputting the fused feature vector and the synthetic fused feature vector into a training dataset for training a virus classification and recognition model. During training, a focus loss function is used to calculate the classification loss value, and the learning rate parameter and network layer number parameter are dynamically adjusted according to the network adjustment coefficient.

[0005] Among them, the fractional differential spectral enhancement processing is based on the fractional calculus theory to adaptively enhance the spectral reflectance curve. The order α of the fractional derivative is a real number between 0 and 1, and the Grünwald-Letnikov definition is used to calculate the α-order fractional derivative at the wavelength.

[0006] The fractional order α is selected by adaptively searching on the training set using the particle swarm optimization algorithm. The objective function is defined as the ratio of the standard deviation of the inter-class spectral distance to the standard deviation of the intra-class spectral distance. Each particle in the particle swarm represents a candidate fractional order.

[0007] Among them, the spectral topology manifold embedding process maps the global geometric structure of high-dimensional spectral data to a low-dimensional feature space based on algebraic topology theory. It treats the spectral reflectance curves of 480 bands as a set of points in 480-dimensional Euclidean space and constructs the Vietoris-Rips complex of the point set.

[0008] Specifically, by gradually increasing the neighborhood radius parameter, a simplex structure with side triangle tetrahedrons is generated from the discrete point set. The Betti number sequence values ​​of the complex under different neighborhood radii are calculated. The zeroth dimension component of the Betti number sequence value represents the number of connected components of the spectral data, and the first dimension component describes the ring topological structure features in the spectral curve.

[0009] The process involves drawing a persistent graph to record the generation and decay of topological features, extracting long-lived topological feature points from the persistent graph to construct topological descriptor vectors, and using kernel principal component analysis to embed the topological descriptor vectors from the high-dimensional topological space into the low-dimensional Euclidean space. The kernel function is a radial basis function.

[0010] The multimodal attention fusion module includes a spectral feature encoding submodule, a thermal infrared feature encoding submodule, and a fluorescence feature encoding submodule. The spectral feature encoding submodule maps the topological embedding feature vector into a spectral intermediate representation vector through a fully connected layer.

[0011] Among them, the thermal infrared feature encoding submodule extracts temperature gradient features from thermal infrared temperature field data through a convolutional neural network and maps them into thermal infrared intermediate representation vectors, while the fluorescence feature encoding submodule extracts temporal decay features from chlorophyll fluorescence dynamics curve data through a long short-term memory network and maps them into fluorescence intermediate representation vectors.

[0012] Specifically, the attention weight matrix between the spectral intermediate representation vector, the thermal infrared intermediate representation vector, and the fluorescence intermediate representation vector is calculated. The attention weight matrix is ​​obtained through a scaled dot product attention mechanism. The weighted intermediate representation vector is obtained by weighting and summing the corresponding row vectors of each intermediate representation vector with the attention weight matrix.

[0013] The spectral generative adversarial network consists of a generator network and a discriminator network. The generator network receives random noise vectors and fused feature vectors of leaves infected with a few types of viruses as inputs, and generates synthetic fused feature vectors through a multi-layer fully connected network and a batch normalization layer.

[0014] The discriminator network receives the fused feature vector or the synthesized fused feature vector as input and outputs the authenticity probability value through a multi-layer fully connected network. The generator network aims to maximize the authenticity probability value of the synthesized fused feature vector by the discriminator network.

[0015] The process involves counting the number of healthy leaf samples and the number of leaf samples infected by various viruses. When the number of leaf samples infected by a certain type of virus is lower than the preset ratio threshold of the number of healthy leaf samples, the fused feature vector corresponding to that type of virus-infected leaf is input into the spectral generative adversarial network.

[0016] The virus classification and identification model uses a focus loss function to calculate the classification loss value. In the focus loss function, the modulation factor is defined as a power function of the predicted probability, and the power exponent ranges from 2.0 to 3.0.

[0017] During training, the network adjustment coefficient is calculated once every preset number of iterations. The network adjustment coefficient is obtained through the network adjustment function. The inputs of the network adjustment function include the overall accuracy of the validation set, the standard deviation of the recall of each class, and the rate of decrease of the loss function corresponding to the current iteration number.

[0018] The calculation of the network adjustment function includes dividing the overall accuracy of the validation set by the preset baseline accuracy to obtain the accuracy normalization value, multiplying the standard deviation of the recall rate of each category by the preset weight coefficient and taking the reciprocal to obtain the recall balance value, and dividing the rate of decrease of the loss function by the preset baseline rate of decrease to obtain the convergence speed normalization value.

[0019] Specifically, the learning rate parameter and the number of network layers are dynamically adjusted based on the network adjustment coefficient. When the network adjustment coefficient is in different ranges, adjustment strategies are adopted respectively, such as keeping the parameters unchanged, reducing the learning rate, increasing convolutional layers, or introducing dropout regularization layers.

[0020] This invention employs fractional-order differential spectral enhancement combined with spectral topological manifold embedding. By continuously adjusting the differential order, it extracts intermediate-level features between the zero-order and first-order derivative spectra, significantly amplifying the subtle changes in chlorophyll content and cell structure damage in the early stages of disease on the spectral curve, thus amplifying the previously submerged differences several times over. Algebraic topology maps the global geometric structure of high-dimensional spectral data to a low-dimensional feature space. Using persistent cohomological eigenvalues ​​and Betti number sequences, it constructs a topological descriptor invariant to translation, scaling, and noise perturbations, effectively capturing the differentiated distribution patterns of different virus types in global morphological features such as peak shifts, valley deepenings, and curvature changes in the spectral curve. Simultaneously, it integrates multimodal data from thermal infrared temperature fields and chlorophyll fluorescence dynamics using an attention mechanism to comprehensively determine disease types from multiple physiological and metabolic dimensions. In summary, this invention solves the technical problem mentioned in the background art of accurately identifying and distinguishing different types of tobacco viruses in the early stages of disease. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 A schematic diagram of the experimental setup.

[0023] Figure 3 The image shows the feature wavelength distribution extracted by four algorithms based on the blade scale. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0025] like Figure 1 The diagram shown is a flowchart of a method for constructing a tobacco virus classification model based on tobacco leaf hyperspectral imaging, provided by this invention. This method includes the following steps:

[0026] S01. Collect hyperspectral reflectance data of tobacco leaves with a wavelength range of 380nm to 1000nm, perform Savitzky-Golay filtering to denoise the hyperspectral reflectance data of tobacco leaves to obtain denoised hyperspectral reflectance data, and perform fractional-order differential spectral enhancement processing on the denoised hyperspectral reflectance data to obtain fractional-order differential enhanced spectral data.

[0027] S02. Perform spectral topological manifold embedding processing on the fractional-order differential enhanced spectral data, calculate the persistent cohomology eigenvalues ​​and Betti number sequence values ​​of the fractional-order differential enhanced spectral data, and reduce the dimensionality of the persistent cohomology eigenvalues ​​and the Betti number sequence values ​​through kernel principal component analysis to obtain the topological embedding feature vector.

[0028] S03. Collect thermal infrared temperature field data and chlorophyll fluorescence dynamics curve data corresponding to the tobacco leaf, and input the topology embedding feature vector, thermal infrared temperature field data and chlorophyll fluorescence dynamics curve data into the multimodal attention fusion module. The multimodal attention fusion module outputs the fused feature vector.

[0029] S04. Count the number of healthy leaf samples and the number of leaf samples infected by various viruses. When the number of leaf samples infected by a certain type of virus is lower than the preset ratio threshold of the number of healthy leaf samples, input the fusion feature vector corresponding to the leaf infected by the certain type of virus into the spectral generation adversarial network. The spectral generation adversarial network outputs the synthesized fusion feature vector.

[0030] S05. The fused feature vector and the synthesized fused feature vector are combined to form a training dataset. The training dataset is input into the virus classification and recognition model for training. The virus classification and recognition model uses the focus loss function to calculate the classification loss value. During the training process, the network adjustment coefficient is calculated once every preset number of iterations. The learning rate parameter and network layer parameter of the virus classification and recognition model are dynamically adjusted according to the network adjustment coefficient.

[0031] S06. Optionally, the method further includes collecting the hyperspectral reflectance data of the tobacco leaves to be tested, the thermal infrared temperature field data, and the chlorophyll fluorescence dynamics curve data of the tobacco leaves to be tested, processing the tobacco leaves to be tested according to steps S01 to S03 to obtain the fused feature vector to be tested, inputting the fused feature vector to be tested into the trained virus classification and recognition model, and the virus classification and recognition model outputs the virus type classification result and the disease severity level result.

[0032] The fractional-order differential spectral enhancement processing adaptively enhances the spectral reflectance curve based on fractional-order calculus theory. The order α of the fractional derivative ranges from 0 to 1, and the fractional derivative extracts intermediate-level feature information between the zero-order spectrum and the first-order derivative spectrum by continuously adjusting the derivative order. The Grünwald-Letnikov definition is used to calculate the α-order fractional derivative at wavelength λ, expanding the spectral curve into an infinite series in the time domain, and obtaining the fractional derivative result through convolution. The fractional order α is selected using a particle swarm optimization algorithm for adaptive search on the training set. The objective function is defined as the ratio of the standard deviation of the inter-class spectral distance to the standard deviation of the intra-class spectral distance. Each particle in the particle swarm represents a candidate fractional order, and the optimal α value that maximizes inter-class separability is found by iteratively updating the particle position and velocity. The fractional order exhibits a differentiated distribution pattern for different virus types and severity levels. Fractional-order differential spectral enhancement significantly amplifies the subtle changes in chlorophyll content and cell structural damage in the early stages of disease on the spectral curve. The previously submerged spectral differences, masked by instrument noise, are enhanced to several times the level of the original signal after fractional-order enhancement. By continuously adjusting the filtering characteristics of the differential operator in the frequency domain, fractional-order differential spectral enhancement achieves adaptive spectral feature extraction for different disease types and severity levels. The multi-scale feature space constructed by fractional-order differential spectral enhancement more comprehensively characterizes the biochemical metabolic disorders caused by viral infection, significantly improving the sensitivity of early disease identification and the accuracy of different virus types.

[0033] The spectral topological manifold embedding process, based on algebraic topology theory, maps the global geometric structure of high-dimensional spectral data to a low-dimensional feature space. The spectral reflectance curves of 480 bands are considered as a set of points in a 480-dimensional Euclidean space. A Vietoris-Rips complex of the point set is constructed, and edge, triangle, and tetrahedral simplex structures are generated by progressively increasing the neighborhood radius parameter. The Betti number sequence values ​​of the complex under different neighborhood radii are calculated. The zeroth dimension of the Betti number sequence value represents the number of connected components in the spectral data, while the first dimension characterizes the annular topological structure features in the spectral curve. A persistent map is plotted to record the generation and disappearance of topological features. Longer-lasting topological features correspond to the essential geometric shape of the spectral curve, while shorter-lasting topological features represent noise perturbations. Long-lived topological feature points in the persistent map are extracted to construct topological descriptor vectors, which are invariant to translation, scaling, and small-amplitude noise perturbations. Kernel principal component analysis (KPCA) is used to embed the topological descriptor vectors from a high-dimensional topological space into a low-dimensional Euclidean space. Radial basis functions are chosen as the kernel function, and the principal component directions are determined by calculating the eigenvalue decomposition of the kernel matrix between the topological descriptors. The dimension of the topologically embedded feature vectors is set to 5% to 10% of the original spectral bands, achieving dimensionality compression while preserving key topological information. Spectral topological manifold embedding fundamentally overcomes the limitations of traditional feature extraction methods that only focus on point-by-point numerical changes in the spectral curve. It captures the global morphological features and multi-scale geometric structure of the spectral curve through topological invariants. Viral infection causing damage to leaf tissue structure manifests in the spectral curve as peak shifts, deepening valleys, and curvature changes within the spectral band range. These local morphological changes are reflected in the topological space as differentiated distribution patterns of the persistent homology eigenvalues. The spectral curves of healthy leaves exhibit a smooth single-peak structure in the visible to near-infrared bands, corresponding to simple connected domains in the topological space. However, the spectral curves of diseased leaves show multi-peak oscillations due to chlorophyll degradation and anthocyanin accumulation, generating additional ring structures in the topological space and causing significant jumps in the Betti number sequence values. Topological features are far more robust to environmental factors such as light variations and leaf angles than absolute numerical features of spectral reflectance, because topological invariants depend only on the relative geometric relationships of the spectral curves rather than their numerical magnitudes, effectively suppressing measurement noise interference in field environments. By decoupling the high-dimensional manifold structure of spectral data into low-dimensional topological features, the spectral topological manifold embedding process preferentially preserves discriminative information that distinguishes different virus types during dimensionality compression, avoiding the category boundary ambiguity problem caused by traditional linear dimensionality reduction methods.

[0034] The preset ratio threshold is between 0.5 and 0.7. The preset iteration count is between 50 and 100. The network adjustment coefficient is obtained through a network adjustment function, which is used to dynamically adjust the model parameters according to the classification performance of the virus classification and identification model on the validation set. The inputs include the overall accuracy of the validation set corresponding to the current iteration count, the standard deviation of the recall rate for each category, and the rate of decrease of the loss function. The output is the network adjustment coefficient. The calculation steps of the network adjustment function are as follows: divide the overall accuracy of the validation set by the preset baseline accuracy to obtain the accuracy normalized value; multiply the standard deviation of the recall rate for each category by the preset weight coefficient and take the reciprocal to obtain the recall balance value; divide the rate of decrease of the loss function by the preset baseline rate of decrease to obtain the convergence speed normalized value; and sum the accuracy normalized value, the recall balance value, and the convergence speed normalized value according to the preset weight ratio to obtain the network adjustment coefficient; when the network adjustment coefficient a∈[ When the network adjustment coefficient a ∈ [0.8, 1.0), the learning rate parameter and the number of network layers remain unchanged. When the network adjustment coefficient a ∈ [0.6, 0.8), the learning rate parameter is reduced to 0.7 times the current value. When the network adjustment coefficient a ∈ [0.4, 0.6), the learning rate parameter is reduced to 0.5 times the current value and an additional convolutional layer is added. When the network adjustment coefficient a < 0.4, the learning rate parameter is reduced to 0.3 times the current value and two convolutional layers are added, while a dropout regularization layer is introduced.

[0035] The preset baseline accuracy is set to 0.75. The preset weight coefficient is set to 2.0 to 3.0. The preset baseline descent rate is set to the average descent rate of the loss function over the first 10 iterations of training. The preset weight ratio is: 0.5 for the normalized accuracy value, 0.3 for the balanced recall value, and 0.2 for the normalized convergence rate value.

[0036] The multimodal attention fusion module includes a spectral feature encoding submodule, a thermal infrared feature encoding submodule, and a fluorescence feature encoding submodule. The spectral feature encoding submodule maps the topological embedding feature vector to a spectral intermediate representation vector through a fully connected layer. The thermal infrared feature encoding submodule extracts temperature gradient features from the thermal infrared temperature field data through a convolutional neural network and then maps them to a thermal infrared intermediate representation vector. The fluorescence feature encoding submodule extracts temporal decay features from the chlorophyll fluorescence kinetic curve data through a long short-term memory network and then maps them to a fluorescence intermediate representation vector. The attention weight matrix among the spectral intermediate representation vector, the thermal infrared intermediate representation vector, and the fluorescence intermediate representation vector is calculated. The attention weight matrix is ​​obtained through a scaled dot product attention mechanism. The weighted spectral intermediate representation vector is obtained by weighted summation of the spectral intermediate representation vector and the first row vector of the attention weight matrix. The weighted thermal infrared intermediate representation vector is obtained by weighted summation of the thermal infrared intermediate representation vector and the second row vector of the attention weight matrix. The weighted fluorescence intermediate representation vector is obtained by weighted summation of the fluorescence intermediate representation vector and the third row vector of the attention weight matrix. The weighted spectral intermediate representation vector, the weighted thermal infrared intermediate representation vector, and the weighted fluorescence intermediate representation vector are concatenated and then dimensionality-reduced through a fully connected layer to obtain the fused feature vector.

[0037] The spectral generative adversarial network consists of a generator network and a discriminator network. The generator network receives a random noise vector and the fused feature vector of a minority virus-infected leaf as input, and generates the synthetic fused feature vector through a multi-layer fully connected network and a batch normalization layer. The discriminator network receives either the fused feature vector or the synthetic fused feature vector as input, and outputs a authenticity probability value through a multi-layer fully connected network. During training, the generator network aims to maximize the authenticity probability value of the synthetic fused feature vector for the discriminator network, while the discriminator network aims to maximize the authenticity probability value of the true fused feature vector while minimizing the authenticity probability value of the synthetic fused feature vector.

[0038] The structure of the virus classification and identification model is as follows: the input layer receives the fused feature vector; the first convolutional layer contains 64 convolutional kernels to extract features and output the first convolutional feature; the first batch of normalization layers normalizes the first convolutional feature and outputs the first normalized feature; the first activation layer uses the ReLU activation function to activate the first normalized feature and outputs the first activation feature; the second convolutional layer contains 128 convolutional kernels to extract deep features from the first activation feature and output the second convolutional feature; the second batch of normalization layers normalizes the second convolutional feature and outputs the second normalized feature; the second activation layer uses the ReLU activation function to activate the second normalized feature and outputs the second activation feature; the fully connected layer maps the second activation feature to a joint probability distribution of virus type and severity level; and the output layer uses the softmax function to output the classification probability.

[0039] The steps for establishing the training dataset for the virus classification and identification model include: collecting healthy tobacco leaf samples and tobacco leaf samples infected with CMV virus, PVY virus, and TMV virus; labeling each tobacco leaf sample with a disease severity level according to the number of days after inoculation: 0 to 2 days as level 0, 3 to 4 days as level 1, 5 to 6 days as level 2, 7 to 8 days as level 3, 9 to 10 days as level 4, 11 to 12 days as level 5, 13 to 14 days as level 6, and 15 days and beyond as level 7; processing all the tobacco leaf samples using steps S01 to S04 to obtain the fused feature vector and the synthesized fused feature vector; and combining the fused feature vector and the synthesized fused feature vector with the corresponding virus type label and severity level label to form the training dataset.

[0040] The steps for training the virus classification and recognition model include: dividing the training dataset into a training subset and a validation subset in an 8:2 ratio; training the model using a five-fold cross-validation method; randomly sampling a batch of samples from the training subset and inputting them into the virus classification and recognition model in each training iteration; calculating the classification loss value calculated by the focus loss function between the model output and the true label; updating the model parameters using the Adam optimizer based on the gradient of the classification loss value; evaluating the model performance on the validation subset and dynamically adjusting the learning rate parameter and the number of network layers parameter according to the network adjustment function; and stopping training when the accuracy of the validation subset no longer improves for 10 consecutive iterations.

[0041] In the focus loss function, the modulation factor is defined as a power function of the predicted probability, with the power exponent ranging from 2.0 to 3.0. For hard-to-classify samples with low predicted probabilities, the modulation factor is close to 1 to maintain a large loss contribution, while for easy-to-classify samples with high predicted probabilities, the modulation factor is close to 0 to reduce the loss contribution.

[0042] The persistent homology eigenvalues ​​are obtained by calculating the homology groups of the simple complex under different neighborhood radius parameters. The generation radius and the disappearance radius of each topological feature are recorded, and the difference between the generation radius and the disappearance radius is defined as the persistence value. The zeroth dimension component of the Betti number sequence value represents the number of connected components, and the first dimension component represents the number of loop structures.

[0043] The dimension of the topological embedding feature vector is adaptively determined based on the cumulative variance contribution rate of the kernel principal component analysis, and the dimension range is 24 to 48. The dimensions of the spectral intermediate characterization vector, the thermal infrared intermediate characterization vector, and the fluorescence intermediate characterization vector are all set to 128 to 256. The dimension of the fused feature vector is set to 192 to 384.

[0044] The generator network comprises four fully connected layers, with each layer having 128, 256, and 512 neurons respectively, corresponding to the dimension of the fused feature vector. The discriminator network comprises three fully connected layers, with each layer having 256 and 1 neurons respectively, corresponding to the dimension of the fused feature vector. The spectral generative adversarial network is trained for 2000 to 5000 iterations, with each iteration updating the discriminator network five times and then the generator network once.

[0045] The random noise vector follows a standard normal distribution with dimensions ranging from 64 to 128. The initial learning rate of the Adam optimizer is set to 0.001 to 0.01, the momentum parameter is set to 0.9, and the second-order momentum parameter is set to 0.999.

[0046] The thermal infrared temperature field data were acquired using a thermal infrared imager with a wavelength response range of 8 μm to 14 μm, a spatial resolution better than 1 mm, and a temperature measurement accuracy better than 0.1 °C. The chlorophyll fluorescence kinetic curve data were acquired using a modulated fluorescence instrument, measuring the changes in photochemical and non-photochemical quenching coefficients over time.

[0047] The Savitzky-Golay filter has a window size of 11 to 21 bands, and the polynomial fitting order is set to 2 to 4. In the kernel principal component analysis, the kernel parameters of the radial basis functions are optimized and determined on the training subset using a grid search method. The number of principal components retained is the minimum number of principal components corresponding to a cumulative variance contribution rate of 95%.

[0048] The virus type classification results include four categories: healthy, CMV virus infection, PVY virus infection, and TMV virus infection. The disease severity level results include eight levels from level 0 to level 7. The training dataset contains 300 to 400 healthy tobacco leaf samples, 400 to 500 CMV virus-infected tobacco leaf samples, 500 to 600 PVY virus-infected tobacco leaf samples, and 450 to 550 TMV virus-infected tobacco leaf samples.

[0049] The dropout probability of the dropout regularization layer is set to 0.3 to 0.5. The scaling factor in the scaled dot product attention mechanism is set to the reciprocal of the square root of the dimension of the intermediate spectral representation vector. The convolutional neural network contains two convolutional layers and one pooling layer. The long short-term memory network contains two LSTM layers, with each layer containing 64 to 128 hidden units.

[0050] The acquisition of the hyperspectral reflectance data of the tobacco leaves was completed using a hyperspectral imager with a spectral resolution better than 1.3 nm, a maximum field of view of 31.7 degrees, and a maximum aperture of F2.6. The hyperspectral reflectance data of the tobacco leaves includes 480 bands.

[0051] The topology descriptor vector is constructed by extracting topological feature points in the persistent graph whose persistence values ​​are greater than a preset persistence threshold, which is set to 1.5 times the median of all persistence values. The neighborhood radius parameter ranges from 0.01 to 1.0, with a step size of 0.01.

[0052] The inter-class spectral distance standard deviation is defined as the standard deviation of the spectral Euclidean distance between samples of different virus types, and the intra-class spectral distance standard deviation is defined as the standard deviation of the spectral Euclidean distance within samples of the same virus type. In the particle swarm optimization algorithm, the number of particles is set to 30 to 50, the maximum number of iterations is set to 100 to 200, the inertia weight is set to 0.6 to 0.9, and the learning factor is set to 1.5 to 2.5.

[0053] Alternatively, the present invention also provides a method for forming a tobacco virus classification and identification system by means of a computer, wherein the computer is provided with a storage medium storing program instructions, and the program instructions execute the above-described method when running in the computer.

[0054] The specific implementation methods of the above steps are described in detail below.

[0055] The specific implementation of step S01 involves data acquisition and preprocessing of tobacco leaves using a hyperspectral imager. First, the hyperspectral imager is fixed 60cm above the leaf, and the light source angle is adjusted to avoid specular reflection. Raw reflectance data for 480 bands within the wavelength range of 380nm to 1000nm is acquired, with a spectral resolution better than 1.3nm for each band. During the acquisition process, stable indoor lighting is maintained, the temperature is controlled between 20℃ and 25℃, and the relative humidity is kept between 55% and 65%. The acquired raw hyperspectral reflectance data is then subjected to Savitzky-Golay filtering. The filtering window size is set to 15 bands, and the polynomial fitting order is set to 3. This filtering algorithm is based on the least squares principle, performing polynomial fitting on the spectral data within a local band range to remove high-frequency noise while preserving the original morphological characteristics of the spectral curve. The filtered data is then denoised hyperspectral reflectance data. Fractional derivative enhancement was performed on the denoised hyperspectral reflectance data. The Grünwald-Letnikov definition was used to calculate the fractional derivative at each wavelength. The fractional order α was determined by searching the training sample set using a particle swarm optimization algorithm. The particle swarm size was set to 40, the maximum number of iterations was set to 150, the initial value of the inertia weight was set to 0.9 and then linearly decreased to 0.4, and the learning factor was set to 2.0. The objective function was defined as maximizing the ratio of the standard deviation of the inter-class spectral distance to the standard deviation of the intra-class spectral distance. The optimal fractional order that maximizes the spectral difference between healthy leaves and diseased leaves was found through iteration. Typical optimization results show that the optimal fractional order is concentrated in the range of 0.3 to 0.7.

[0056] The specific implementation of step S02 involves performing topological manifold embedding dimensionality reduction processing on the fractional-differential enhanced spectral data. First, the 480-dimensional spectral data point set is constructed as a Vietoris-Rips complex. The neighborhood radius parameter starts at 0.01 and gradually increases to 1.0 in steps of 0.01. For each neighborhood radius, the Betti number sequence of the complex is calculated. The zero-dimensional Betti number represents the number of connected components in the spectral data, and the first-dimensional Betti number represents the number of cyclic topological structures. The neighborhood radius where each topological feature appears and the neighborhood radius where it disappears are recorded; the difference between the two is defined as the persistence value. Topological features with persistence values ​​greater than 1.5 times the median are considered stable features. The generation and disappearance radii of these stable features are extracted as persistent homology feature values. A persistence map is plotted to visualize the topological features on a two-dimensional plane. The horizontal axis represents the generation radius, and the vertical axis represents the disappearance radius. Points far from the diagonal represent persistent, essential topological features, while points close to the diagonal represent less persistent, noisy topological features. The persistent homology eigenvalues ​​and Betti number sequence values ​​are used to form a topological descriptor vector with a dimension of approximately 80 to 120. Kernel principal component analysis is performed on this vector to reduce its dimension. The kernel function is a radial basis function, and the kernel parameters are optimized in the range of 0.01 to 10 through grid search. The number of principal components corresponding to a cumulative variance contribution rate of 95% is retained. The resulting topological embedding feature vector has a dimension of 24 to 48. This feature vector inherits the global geometric morphology information of the spectral curve and is robust to noise perturbations.

[0057] The specific implementation of step S03 involves fusing multi-source data. First, a thermal infrared imager is used to collect the temperature field distribution on the leaf surface. The imager has a wavelength response range of 8μm to 14μm, a spatial resolution of 1mm, and a temperature measurement accuracy of 0.1℃. Temperature gradients are calculated on the collected thermal infrared images. These temperature gradients reflect the differences in metabolic activity in different areas of the leaf. Healthy leaves exhibit uniform temperature distribution, while diseased leaves show localized temperature increases or decreases in the lesion areas. Chlorophyll fluorescence kinetics data are collected using a modulated fluorescence instrument. The temporal changes in photochemical quenching coefficients and non-photochemical quenching coefficients during the light excitation process after dark adaptation are measured. The fluorescence kinetics curves reflect the efficiency of the electron transport chain and the heat dissipation capacity of the leaf's photosynthetic system. The topological embedding feature vector is input into the spectral feature encoding submodule. This submodule contains two fully connected layers with 128 and 256 neurons per layer, respectively. The ReLU function is used as the activation function, and the output is a spectral intermediate representation vector. Thermal infrared temperature field data is input into the thermal infrared feature encoding submodule, which contains two convolutional layers and one max-pooling layer. The convolutional kernel size is 3x3, the stride is 1, and the number of output channels is 32 and 64, respectively, outputting thermal infrared intermediate representation vectors. Chlorophyll fluorescence dynamics curve data is input into the fluorescence feature encoding submodule, which contains two long short-term memory network layers, each with 96 hidden units, outputting fluorescence intermediate representation vectors. The attention weight matrix among the three intermediate representation vectors is calculated using a scaled dot product attention mechanism. The scaling factor is set to the reciprocal of the square root of the dimension of the intermediate representation vector. The attention weights are obtained by normalization using the softmax function. These weights reflect the contribution of different modalities of data to virus classification. The three intermediate representation vectors are weighted and summed with their corresponding attention weights to obtain a weighted intermediate representation vector, which is then concatenated. The concatenated vector has a dimension of 768 to 1152. The dimension is then reduced to 192 to 384 through a fully connected layer to obtain the fused feature vector.

[0058] The specific implementation of step S04 addresses the imbalanced training data problem. First, the number of healthy leaf samples and the number of leaf samples infected with CMV, PVY, and TMV viruses in the training set are statistically analyzed. When the ratio of the number of infected leaf samples to the number of healthy leaf samples is less than 0.6, this category is determined to be a minority class, requiring data augmentation. The fused feature vector of the minority class samples is input into a spectral generative adversarial network (GAN). The generator network receives a 64-dimensional random noise vector following a standard normal distribution and the minority class fused feature vector as input. The feature dimension is progressively expanded through four fully connected layers: the first layer has 128 neurons, the second 256, the third 512, and the fourth layer represents the fused feature vector dimension. Each layer is followed by a batch normalization layer and a Leaky ReLU activation layer, outputting a synthesized fused feature vector. The discriminator network receives either the real or synthesized fused feature vector as input and progressively compresses the feature dimension through three fully connected layers, outputting a probability value between 0 and 1 indicating realism. The training process employs an alternating optimization strategy. In each iteration, the generator parameters are fixed and the discriminator network is updated 5 times. The discriminator loss function consists of the negative log-likelihood of the real samples and the negative log-likelihood of the synthetic samples. Then, the discriminator parameters are fixed and the generator network is updated once. The generator loss function is the negative logarithm of the probability of the discriminator outputting the realness of the synthetic samples. The total number of training iterations is 3000. The initial learning rate is set to 0.0002, and it decays to 0.8 times the original value every 500 iterations. After training, the generator is used to generate synthetic fusion feature vectors to supplement the training dataset.

[0059] The specific implementation of step S05 involves training a virus classification and recognition model. First, the training dataset containing fused feature vectors and synthetic fused feature vectors is randomly divided into training and validation subsets at an 8:2 ratio. A five-fold cross-validation method is used, dividing the training subset into five equal parts. Four parts are selected for training and one for validation each time, repeating this process five times to ensure each data set is used as a validation set at least once. The input layer of the virus classification and recognition model receives the fused feature vectors. The first convolutional layer contains 64 3x3 convolutional kernels with a stride of 1 and same padding. After outputting the first convolutional feature, it is normalized through a first batch of normalization layers. The normalization parameters include mean and variance. The first activation layer uses the ReLU activation function. The second convolutional layer contains 128 convolutional kernels, and subsequent processing is the same as the first convolutional layer. The fully connected layer flattens the second activation feature and maps it to a joint probability distribution vector of virus type and severity level. The output layer normalizes the probability vector using the softmax function. The model training employs a focus loss function, which introduces a modulation factor on top of the cross-entropy loss. The modulation factor is defined as 1 minus a power of the predicted probability, with the exponent set to 2.5. This mechanism ensures that hard-to-classify samples with low predicted probabilities receive greater loss weights, while easy-to-classify samples with high predicted probabilities receive smaller loss weights, effectively mitigating class imbalance. The Adam optimizer is used to update model parameters, with an initial learning rate of 0.005, a momentum parameter of 0.9, a second-order momentum parameter of 0.999, and a batch size of 32. During training, model performance is evaluated on a validation subset every 50 iterations. The overall accuracy of the validation set, the standard deviation of recall for each class, and the rate of decrease of the loss function are calculated. The overall accuracy of the validation set is divided by 0.75 to obtain the normalized accuracy value. The standard deviation of recall for each class is multiplied by 2.5 and the reciprocal is taken to obtain the recall balance value. The rate of decrease of the loss function is divided by the average rate of decrease of the loss function in the first 10 iterations of training to obtain the normalized convergence rate value. The three are weighted and summed according to the weight ratios of 0.5, 0.3, and 0.2 to obtain the network adjustment coefficient. The learning rate parameter and the number of network layers are dynamically adjusted according to the range of the coefficient. Training is stopped early when the accuracy of the validation subset no longer improves for 10 consecutive iterations.

[0060] The specific implementation of step S06 involves classifying and identifying viruses in the tobacco leaves to be tested. First, hyperspectral reflectance data, thermal infrared temperature field data, and chlorophyll fluorescence dynamics curve data of the tobacco leaves to be tested are collected. The hyperspectral reflectance data is subjected to Savitzky-Golay filtering for noise reduction and fractional derivative enhancement processing according to the method in step S01. The fractional derivative enhanced spectral data is subjected to topological manifold embedding dimensionality reduction according to the method in step S02 to obtain a topological embedding feature vector. The topological embedding feature vector, thermal infrared temperature field data, and chlorophyll fluorescence dynamics curve data are input into the multimodal attention fusion module according to the method in step S03 to obtain the fusion feature vector to be tested. The fused feature vector to be tested is input into the trained virus classification and recognition model. The model performs forward propagation calculations layer by layer. The first convolutional layer extracts shallow features, the second convolutional layer extracts deep features, and the fully connected layer maps the features to a joint probability distribution of virus type and severity level. The output layer outputs the classification probability of each category through the softmax function, selects the category with the highest probability as the virus type classification result, and outputs the category with the second highest probability as the suboptimal result for reference. The severity level result is selected from eight levels from level 0 to level 7, and the level with the highest probability is output.

[0061] It should be noted that one of the key technical ideas of this invention is fractional-order differential spectral enhancement processing technology. This technology overcomes the limitation of traditional integer-order derivatives, which can only extract features at a fixed scale. By continuously adjusting the derivative order between 0 and 1, a multi-scale feature space is constructed between the zero-order spectrum and the first-order derivative spectrum. This adaptively extracts the most discriminative intermediate-level feature information based on the differences in spectral response characteristics of different virus types and severity levels. Fractional-order differentiation, by adjusting the filtering characteristics of the differential operator in the frequency domain, amplifies the signal of subtle biochemical changes in the early stages of disease, while suppressing the influence of instrument noise and environmental interference. This significantly enhances subtle spectral differences that were originally submerged in noise, effectively improving the sensitivity and accuracy of early disease identification. Compared to traditional spectral preprocessing methods that only perform smoothing and denoising or first-order differentiation, the continuous feature space constructed by fractional-order differentiation more comprehensively characterizes the physiological and biochemical processes caused by virus infection, such as chlorophyll degradation, cell structure damage, and water metabolism disorders.

[0062] The second key technical approach of this invention is spectral topological manifold embedding dimensionality reduction technology. Based on algebraic topology theory, this technology maps the global geometric structure of high-dimensional spectral data to a low-dimensional feature space. By calculating persistent homology features and Betti number sequences, it captures the topological invariants of the spectral curves. These topological features reflect the essential geometric shape of the spectral curves rather than point-by-point numerical changes, exhibiting natural robustness to translation, scaling, and noise perturbations. Healthy leaves and leaves infected with different viruses exhibit differentiated peak-valley distribution patterns in the visible to near-infrared spectral bands. These morphological differences are manifested in the topological space as significant variations in the number of connected components and the number of ring structures. By separating essential topological features from noisy topological features through persistent spectral mapping, the extracted topological descriptor vectors effectively characterize the global morphological variations of the spectral curves caused by viral infection. Compared to traditional principal component analysis or linear discriminant analysis, which only focuses on linear transformations of spectral data, topological manifold embedding technology captures the nonlinear manifold structure of spectral data. While significantly reducing feature dimensionality, it prioritizes the preservation of discriminative information distinguishing different virus types, avoiding the curse of dimensionality and blurred category boundaries.

[0063] The third key technical approach of this invention is data augmentation based on generative adversarial networks (GANs). This technique addresses the imbalance in the number of samples of different virus types in the training data by learning the feature distribution patterns of minority class samples through an adversarial training mechanism, generating high-fidelity synthetic samples to supplement the training dataset. The generator network explores the latent feature manifold of minority class samples in a random noise space, while the discriminator network learns the feature boundaries of real and synthetic samples. The alternating optimization of both networks gradually enables the generator to approximate the statistical distribution characteristics of real samples. Compared to traditional synthetic minority oversampling techniques that only perform linear interpolation between existing minority class samples, GANs can explore unknown regions in the sample space. The generated synthetic samples have richer feature diversity and more realistic spectral details of diseases, effectively expanding the feature coverage of minority class samples. Combined with the focus loss function's automatic focusing mechanism for difficult-to-classify samples, this technique significantly improves the model's ability to identify minority virus types and avoids the problem of model predictions being biased towards the majority class.

[0064] The synergistic effect of the three key technical approaches described above constitutes the core technical advantage of this invention. Fractional-order differential spectral enhancement processing improves the discriminative power of spectral features at the source, providing high-quality input data for subsequent topological manifold embedding. Topological manifold embedding technology achieves efficient dimensionality reduction while preserving essential spectral morphological features, reducing computational complexity for multimodal fusion and model training. Generative adversarial network data augmentation technology solves the problem of imbalanced training data, providing sufficient sample support for the model to learn minority virus characteristics. These three technologies form a complete technical chain from feature enhancement to feature extraction to model training. The technological innovations at each stage support and promote each other, jointly achieving high-precision early identification and accurate classification of multiple types of tobacco viral diseases. Compared with existing technologies that rely solely on raw spectral data or single dimensionality reduction methods, the multi-level feature extraction and intelligent sample enhancement system constructed in this invention has achieved significant improvements in recognition sensitivity, classification accuracy, and model robustness.

[0065] It should be noted that this invention also solves the following technical problem: In the practical application of tobacco virus detection, due to the significant differences in the incidence rates of different virus types, the collected sample data exhibits a severe class imbalance distribution. The number of samples infected with minority virus types is far lower than that of healthy samples and common virus samples. Traditional classification models tend to optimize the classification performance of majority class samples during training while ignoring minority class samples, resulting in a low recall rate for rare virus types. This invention statistically analyzes the number of samples of each class and sets a preset proportion threshold. When the number of samples infected with a certain type of virus is lower than this threshold, the fused feature vector of that type of sample is input into a spectral generative adversarial network. The generator network receives random noise vectors and minority class real sample features and generates synthetic feature vectors through a multi-layer fully connected network. The discriminator network learns to distinguish between real and synthetic features. Through adversarial training between the generator and the discriminator, the feature distribution of synthetic samples approximates the distribution of real samples, effectively expanding the number of minority class samples. At the same time, a focus loss function is used to assign greater loss weights to difficult-to-classify samples through modulation factors, and the network parameters are dynamically adjusted according to the standard deviation of the recall rate of each class on the validation set to ensure that the model maintains balanced recognition performance for all virus types.

[0066] Specifically, the principle of this invention is as follows: This invention constructs a multi-scale feature space through fractional-order differential enhancement processing to more comprehensively characterize the biochemical metabolic disorder process caused by viral infection. Fractional-order differential processing achieves adaptive feature extraction for different disease severity levels by continuously adjusting the filtering characteristics of the differential operator in the frequency domain. The particle swarm optimization algorithm adaptively searches for the optimal fractional order that maximizes inter-class separability, exhibiting differentiated optimal order distribution patterns for different virus types. Topological manifold embedding processing overcomes the limitations of traditional methods that only focus on point-by-point numerical changes. The tissue structural damage caused by viral infection manifests as a differentiated distribution of persistent homology eigenvalues ​​in the topological space. The spectral curves of healthy leaves correspond to simple connected domains, while diseased leaves exhibit multi-peak oscillations due to chlorophyll degradation and anthocyanin accumulation, generating additional ring structures in the topological space, causing significant jumps in the Betti number sequence values. Topological invariants depend only on relative geometric relationships rather than numerical magnitude, thus exhibiting strong robustness to environmental factors. During dimensionality compression, discriminative information distinguishing different virus types is preferentially preserved.

[0067] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0068] The specific implementation of step S01 is as follows: Hyperspectral reflectance data of tobacco leaves with a wavelength range of 380 nm to 1000 nm is collected, and the hyperspectral reflectance data of tobacco leaves contains 480 bands. The hyperspectral reflectance data of tobacco leaves is then subjected to Savitzky-Gore filtering to obtain denoised hyperspectral reflectance data. The window size of the Savitzky-Gore filter is set to 11 to 21 bands, and the polynomial fitting order is set to 2 to 4. The denoised hyperspectral reflectance data is then subjected to fractional-derivative spectral enhancement processing to obtain fractional-derivative enhanced spectral data, where the order of the fractional derivative is... The wavelength is calculated using a real number ranging from 0 to 1, and the wavelength is determined by the Grünwald-Laitnikov definition. place The fractional derivative is expressed by the following formula:

[0069] ;

[0070] In the formula, wavelength place The fractional derivative is dimensionless. wavelength The denoised hyperspectral reflectance at the location is dimensionless. The wavelength is the spectral wavelength, measured in nm, and ranges from 380 nm to 1000 nm. The original fractional derivative value at a reference wavelength of 550 nm is used for normalization. This is the wavelength sampling interval, measured in nm, and is typically 1.3 nm. The exponent is a summation exponent, with values ​​ranging from 0 to infinity. Let be the coefficient of the generalized binomial, denoted as Dimensionless. Fractional order. The selection is performed using the particle swarm optimization algorithm on the training set adaptively. The objective function is defined as follows:

[0071] ;

[0072] In the formula, The objective function value is dimensionless. The standard deviation of inter-class spectral distance is expressed in nm. The standard deviation of the intra-class spectral distance is expressed in nm. The maximum spectral Euclidean distance in the training set, expressed in nm, is used for normalization to ensure that both the numerator and denominator are dimensionless. The standard deviation of the inter-class spectral distance is defined as the standard deviation of the spectral Euclidean distance between samples of different virus types, and the standard deviation of the intra-class spectral distance is defined as the standard deviation of the spectral Euclidean distance within samples of the same virus type. In the particle swarm optimization algorithm, the number of particles is set to 30 to 50, the maximum number of iterations is set to 100 to 200, the inertia weight is set to 0.6 to 0.9, and the learning factor is set to 1.5 to 2.5.

[0073] The specific implementation of step S02 is as follows: The fractional-order differential enhanced spectral data is subjected to spectral topological manifold embedding processing. The spectral reflectance curves of the 480 bands are regarded as a point set in a 480-dimensional Euclidean space. A Vittoris-Lipps complex of the point set is constructed by gradually increasing the neighborhood radius parameter. The simplex structure is generated by expanding from a discrete point set, wherein the neighborhood radius parameter The value ranges from 0.01 to 1.0, with a step size of 0.01, and is dimensionless. The Betti number sequence values ​​of the complex under different neighborhood radii are calculated, and the zeroth dimension component of the Betti number sequence value is... The number of connected components in the spectral data is dimensionless; the first dimension component... The number of ring-shaped topological structures in the spectral curve is dimensionless. The persistent homology eigenvalues ​​are obtained by calculating the homology groups of the simple complex under different neighborhood radius parameters, and the generation radius of each topological feature is recorded. and extinction radius The persistence value is defined as follows:

[0074] ;

[0075] In the formula, For the first The persistence value of each topological feature is dimensionless; For the first The vanishing radius of a topological feature, dimensionless; For the first The generation radius of each topological feature, dimensionless; This serves as an index for topological features. Topological feature points with persistence values ​​greater than a preset persistence threshold are extracted from the persistent graph to construct topological descriptor vectors. The preset persistence threshold is set to 1.5 times the median of all persistence values. Kernel principal component analysis is used to embed the topological descriptor vectors from the high-dimensional topological space into the low-dimensional Euclidean space. The kernel function chosen is a radial basis function, expressed in the following formula:

[0076] ;

[0077] In the formula, For topological descriptors and The kernel function values ​​between these values ​​are dimensionless. and A topological descriptor vector; and For the index of the topological descriptor; The kernel parameter is dimensionless and is determined by optimization on the training subset using a grid search method; its empirical value is 0.01 to 0.1. The distance is Euclidean, with the same dimensions as the topological descriptor vector. The dimension of the topological embedding feature vector is adaptively determined based on the cumulative variance contribution rate of the kernel principal component analysis. The number of principal components retained is the minimum number of principal components corresponding to a cumulative variance contribution rate of 95%, with a dimension range of 24 to 48.

[0078] The specific implementation of step S03 is as follows: Thermal infrared temperature field data and chlorophyll fluorescence dynamics curve data corresponding to the tobacco leaves are collected. The thermal infrared temperature field data is collected using a thermal infrared imager with a wavelength response range of 8 Å. up to 14 The spatial resolution is better than 1 mm, and the temperature measurement accuracy is better than 0.1 °C. The chlorophyll fluorescence kinetic curve data were acquired using a modulated fluorescence spectrometer, measuring the changes in photochemical and non-photochemical quenching coefficients over time. The topological embedding feature vector, along with the thermal infrared temperature field data and the chlorophyll fluorescence kinetic curve data, were input into a multimodal attention fusion module. This multimodal attention fusion module includes a spectral feature encoding submodule, a thermal infrared feature encoding submodule, and a fluorescence feature encoding submodule. The spectral feature encoding submodule maps the topological embedding feature vector into a spectral intermediate representation vector through a fully connected layer. The thermal infrared feature encoding submodule extracts temperature gradient features from the thermal infrared temperature field data using a convolutional neural network and then maps them into a thermal infrared intermediate representation vector. The fluorescence feature encoding submodule extracts the temporal decay features from the chlorophyll fluorescence kinetic curve data through a long short-term memory network and maps them into a fluorescence intermediate representation vector. The convolutional neural network comprises two convolutional layers and one pooling layer, and the long short-term memory network comprises two long short-term memory layers, with each layer containing 64 to 128 hidden units. The dimensions of the spectral intermediate representation vector, the thermal infrared intermediate representation vector, and the fluorescence intermediate representation vector are all set to 128 to 256. An attention weight matrix is ​​calculated among the spectral intermediate representation vector, the thermal infrared intermediate representation vector, and the fluorescence intermediate representation vector. This attention weight matrix is ​​obtained through a scaled dot product attention mechanism, and the three intermediate representation vectors are denoted as a set. ,in , , The formula for calculating the elements of the attention weight matrix is ​​as follows:

[0079] ;

[0080] In the formula, The first of the attention weight matrix Line 1 Column elements, dimensionless; and The first The and the first One intermediate representation vector; and These are the row index and column index, respectively, with values ​​ranging from 1 to 3; For the first intermediate representation vectors The summation index takes values ​​from 1 to 3; The dimension of the intermediate representation vector of the spectrum; This is the scaling factor; This represents the vector dot product operation. The weighted spectral intermediate representation vector is obtained by weighted summation of the spectral intermediate representation vector and the first row vector of the attention weight matrix. The weighted thermal infrared intermediate representation vector is obtained by weighting and summing the thermal infrared intermediate representation vector with the second row vector of the attention weight matrix. The weighted intermediate fluorescence representation vector is obtained by weighting and summing the intermediate fluorescence representation vector with the third row vector of the attention weight matrix. The fused feature vectors are expressed as follows:

[0081] ;

[0082] In the formula, To fuse feature vectors; This is the weight matrix of the fully connected layer; , , These are the weighted spectral intermediate characterization vector, the weighted thermal infrared intermediate characterization vector, and the weighted fluorescence intermediate characterization vector, respectively. This represents a vector concatenation operation, which joins three weighted intermediate representation vectors column-wise into a longer vector. The dimension of the fused feature vector is set to between 192 and 384.

[0083] The specific implementation of step S04 is as follows: count the number of healthy leaf samples. and the number of leaf samples infected with various viruses When the number of leaf samples infected by a certain type of virus is lower than a preset threshold proportion of the number of healthy leaf samples, the judgment condition is stated as follows:

[0084] ;

[0085] In the formula, For the first Number of leaf samples infected with viroids; Number of healthy leaf samples; The preset ratio threshold is dimensionless and ranges from 0.5 to 0.7. This serves as a virus type index. The fused feature vector corresponding to the leaf infected by the virus of a certain type is input into a spectral generative adversarial network (GAN), which consists of a generator network and a discriminator network. The generator network receives a random noise vector. The fused feature vectors of leaves infected with a few virus-like organisms are used as input, and the generator input is represented as follows: ,in It is a random noise vector. To fuse feature vectors, The vector concatenation operation generates the synthesized fused feature vector through a multi-layer fully connected network and a batch normalization layer. The random noise vector follows a standard normal distribution with dimensions ranging from 64 to 128. The generator network contains four fully connected layers, with the number of neurons in each layer being 128, 256, and 512 respectively, corresponding to the dimension of the fused feature vector. The discriminator network contains three fully connected layers, with the number of neurons in each layer being the dimension of the fused feature vector, 256, and 1 respectively. The spectral generative adversarial network is trained for 2000 to 5000 iterations, with each iteration updating the discriminator network five times and then the generator network once.

[0086] The specific implementation of step S05 is as follows: The fused feature vector and the synthesized fused feature vector are combined to form a training dataset. The training dataset contains 300 to 400 healthy tobacco leaf samples, 400 to 500 tobacco leaf samples infected with cucumber mosaic virus, and infected potato samples... The number of tobacco leaf samples infected with the virus was 500 to 600, and the number of tobacco leaf samples infected with tobacco mosaic virus was 450 to 550. The training dataset was input into the virus classification and recognition model for training. The virus classification and recognition model used a focus loss function to calculate the classification loss value. The focus loss function is described as follows:

[0087] ;

[0088] In the formula, The focal loss value is dimensionless. Total number of categories; For the first The predicted probability of a class is dimensionless and ranges from 0 to 1. For the first The actual label of the class, dimensionless, with a value of 0 or 1; Indexed by category; The modulation factor is dimensionless. The modulation factor is a power-law exponent, dimensionless, and ranges from 2.0 to 3.0. During training, the network adjustment coefficient is calculated once every preset number of iterations, where the preset number of iterations ranges from 50 to 100. The network adjustment coefficient... Obtained through network adjustment functions, as described below:

[0089] ;

[0090] In the formula, The network adjustment coefficients are dimensionless. , , The preset weight ratios are dimensionless and take values ​​of 0.5, 0.3, and 0.2 respectively. To verify the overall accuracy of the set, dimensionless; The preset baseline accuracy is dimensionless and has a value of 0.75. represents the standard deviation of recall rates for each category, which is dimensionless. The preset weighting coefficients are dimensionless and range from 2.0 to 3.0. The rate of decrease of the loss function is calculated using the following formula: ,in and These are the previous and current loss function values, respectively. The initial loss value is used for normalization. The iteration interval is determined by normalization. It is a dimensionless value; The preset baseline descent rate is dimensionless and is taken as the average descent rate of the loss function over the first 10 iterations of training. It is used in conjunction with... The same calculation method; This is the iteration count index. Adjusted according to the network adjustment coefficients. Dynamically adjust the learning rate and network layer number parameters of the virus classification and recognition model when While maintaining the current learning rate parameter and the network layer number parameter unchanged, when When the learning rate parameter is reduced to 0.7 times its current value, When the learning rate parameter is reduced to 0.5 times the current value and a convolutional layer is added, The learning rate parameter is reduced to 0.3 times the current value, and two convolutional layers are added while a dropout regularization layer is introduced. The dropout probability of the dropout regularization layer is set to 0.3 to 0.5. The initial learning rate of the Adam optimizer is set to 0.001 to 0.01, the momentum parameter is set to 0.9, and the second-order momentum parameter is set to 0.999.

[0091] The specific implementation method of step S06 is the same as described above, and will not be repeated in detail here.

[0092] To better understand and implement this invention, a specific application scenario of the invention is provided below as Example 2: To address the practical problems of early identification and severity grading of tobacco viral diseases using the technical solution of this invention, technicians constructed a tobacco virus classification and testing system based on hyperspectral imaging. The test was conducted in a greenhouse environment. CMV virus, PVY virus, and TMV virus were artificially inoculated onto healthy tobacco leaves, respectively. A control group of healthy plants was set up, such as... Figure 2 The diagram shows the overall experimental setup. Technicians used a hyperspectral imager to acquire leaf reflectance data in the wavelength range of 380 nm to 1000 nm, with a spectral resolution of 1.2 nm, obtaining spectral information for a total of 480 bands. Simultaneously, a thermal infrared imager was used to acquire leaf temperature field data, with a wavelength response range of 8 μm to 14 μm, a temperature measurement accuracy of 0.08 °C, and a spatial resolution of 0.8 mm. A modulated fluorescence spectrometer was used to simultaneously record chlorophyll fluorescence kinetic curves, measuring the temporal variations of photochemical and non-photochemical quenching coefficients.

[0093] After data collection, technicians used reverse transcription polymerase chain reaction (RT-PCR) technology for molecular identification and screening of samples. RNA was extracted using Trizol and chloroform, and converted to cDNA via reverse transcription in RT-PCR. The cDNA was then amplified by polymerase chain reaction PCR using Trizol and chloroform. The reagents used were Trizol's EasyScript One-Step gDNA Removal and cDNA Synthesis SuperMix and 2×EasyTaq PCR SuperMix. The templates were amplified using an RT-PCR instrument. Gel electrophoresis was used to sequence the amplified products to verify their sequence accuracy. Sequencing results were compared to exclude non-target samples, and samples meeting the requirements were selected to ensure the accuracy and reliability of subsequent analyses. The amplification primer sequences for each gene are shown in Table 1. Sequencing was performed by Qingdao Weilai Biotechnology Co., Ltd., and the experiment was repeated three times. After sequence alignment, 20 healthy plants, 21 plants infected with CMV, 26 plants infected with PVY, and 25 plants infected with TMV were selected, totaling 645 GB of data. Images of non-target samples are deleted, and the retained image data is used for subsequent screening and feature band analysis.

[0094] Table 1 RT-PCR primer sequences

[0095]

[0096] Plant spectral information is prone to noise bands. Technicians preprocess the spectral data to filter out useful information relevant to the plant. For high-frequency noise in the spectral signal, the Savitzky-Golay convolutional smoothing algorithm has significant advantages in signal denoising, improving spectral continuity while maintaining the fidelity of the original spectral features. For noise bands such as sawtooth patterns on the spectral curve, the SG filtering algorithm can improve the smoothness of the spectral curve while preserving the original spectral information. The SMOTE algorithm is used to address the data imbalance problem, generating new data based on the proximity relationships of minority class samples in the feature space, thus expanding its feature distribution range. The data processing stage uses SpectralAnalysis software developed by Aopu Tiancheng Enterprise for processing and analysis.

[0097] Viral infection of leaves produces mosaic and mottling, altering the overall leaf spectrum. To further investigate the spectral changes caused by the viral disease affecting the internal leaf structure, an imaging hyperspectral analyzer was used to record spectral information of diseased plants during their growth process over time. The reflectance difference between CMV- and TMV-infected plants and PVY-infected plants gradually increased over time, with PVY plants showing a gradual decrease in reflectance compared to the other two diseased plants. Overall, the spectral reflectance of diseased plants gradually increased compared to healthy plants in the visible and red-edge bands, while the difference widened in the near-infrared region. This may be due to viral infection damaging the photosynthetic system of tobacco leaves, reducing the light-capturing capacity of photosynthetic pigments, thus increasing reflectance in the visible light band; the decreased reflectance in the near-infrared region may be related to viral damage to mesophyll cell structure and water metabolism disorders. Abnormalities in the later spectral curves may be due to uneven light intensity distribution across different parts of the image caused by increased plant growth.

[0098] Technicians first performed Savitzky-Golay filtering to denoise the collected hyperspectral reflectance data, setting the window size to 15 bands and the polynomial fitting order to 3. The denoised spectral data then underwent fractional-derivative spectral enhancement, using a particle swarm optimization algorithm to adaptively search for the optimal fractional-order α value. The particle swarm consisted of 40 particles, with a maximum iteration count of 150, an inertia weight of 0.75, and a learning factor of 2.0. After optimization, the optimal fractional-order α was found to be 0.62 for CMV-infected leaves, 0.58 for PVY virus, and 0.67 for TMV virus. The fractional-derivative enhanced spectral data significantly amplified the subtle changes in chlorophyll content in the early stages of disease on the spectral curve, increasing the discriminative power of weak spectral differences previously submerged in instrument noise to 4.3 times that of the original signal.

[0099] When performing spectral topological manifold embedding on fractional-order differential enhanced spectral data, the technicians constructed a Vietoris-Rips complex from the 480-dimensional spectral reflectance curves. The neighborhood radius parameter was gradually increased from 0.01 to 1.0 with a step size of 0.01, and persistent homology eigenvalues ​​and Betti number sequence values ​​were calculated at different neighborhood radii. The spectral curves of healthy leaves exhibited simple connected domains in the topological space, with the zeroth-dimensional Betti number remaining stable at 1 and the first-dimensional Betti number remaining at 0. CMV-infected leaves, due to chlorophyll degradation and anthocyanin accumulation, showed multi-peak oscillations in their spectral curves, generating additional loop structures in the topological space, with the first-dimensional Betti number jumping to 2 at a neighborhood radius of 0.34. Topological feature points with persistence values ​​greater than 1.5 times the median of all persistence values ​​were extracted to construct a topological descriptor vector with a dimension of 156. Kernel principal component analysis was used for dimensionality reduction. The kernel parameter of the radial basis function was determined to be 0.85 through grid search optimization. The 36 principal components corresponding to the cumulative variance contribution rate of 95% were retained, resulting in a 36-dimensional topological embedding feature vector.

[0100] Technicians input topological embedding feature vectors along with thermal infrared temperature field data and chlorophyll fluorescence dynamics curve data into a multimodal attention fusion module. The spectral feature encoding submodule maps the 36-dimensional topological embedding feature vector to a 192-dimensional spectral intermediate representation vector through a fully connected layer. The thermal infrared feature encoding submodule uses a convolutional neural network with two convolutional layers and one pooling layer to extract temperature gradient features, mapping them to a 192-dimensional thermal infrared intermediate representation vector. The fluorescence feature encoding submodule extracts temporal decay features through two LSTM layers, with 96 hidden units per layer, mapping them to a 192-dimensional fluorescence intermediate representation vector. The attention weight matrix between the three intermediate representation vectors is calculated, with a scaling factor set to [value missing]. The reciprocal of the product. After weighted summation using a scaled dot product attention mechanism, the three weighted intermediate representation vectors are concatenated and reduced to 256 dimensions through a fully connected layer to obtain the fused feature vector.

[0101] Of the collected samples, 360 were healthy leaves, 450 were CMV-infected leaves, 540 were PVY-infected leaves, and 500 were TMV-infected leaves. Technicians found that healthy leaves accounted for 19.5% of the total samples, while the number of leaves infected with each virus was more than 1.25 times higher than the number of healthy samples, exceeding the preset proportion threshold of 0.6. Therefore, there was no need to perform generative adversarial network data augmentation on the samples; all fused feature vectors were directly used to form the training dataset. The samples were labeled according to virus type and severity level. The severity level was divided into 8 levels based on the number of days after vaccination: 0-2 days was labeled as level 0, 3-4 days as level 1, 5-6 days as level 2, 7-8 days as level 3, 9-10 days as level 4, 11-12 days as level 5, 13-14 days as level 6, and 15 days and beyond as level 7.

[0102] Technicians used four methods—SPA, GA, CARS, and WOA—to select characteristic bands, such as... Figure 3 The figure shows the distribution of feature wavelengths extracted by four algorithms. The SPA method selected 45 feature bands, mainly distributed in the range of 597nm to 702nm. The GA method selected 16 feature bands, concentrated in the region of 620nm to 690nm. The CARS method selected 114 bands, distributed in the ranges of 422nm to 460nm, 623nm to 652nm, 668nm to 717nm, 766nm to 780nm, and 801nm to 812nm. The WOA method selected 235 bands, uniformly distributed across the entire spectrum. The important wavelengths selected by the four methods highly overlap in the range of 623nm to 702nm, indicating that this band range has a significant discriminative effect on the classification of tobacco viral diseases.

[0103] The training dataset was divided into training and validation subsets in an 8:2 ratio to build a virus classification and recognition model. The model structure includes an input layer, a first convolutional layer with 64 kernels, a first batch of normalization layers, a first ReLU activation layer, a second convolutional layer with 128 kernels, a second batch of normalization layers, a second ReLU activation layer, a fully connected layer, and a softmax output layer. A focus loss function was used to calculate the classification loss value, with a power exponent set to 2.5, maintaining a larger weight for difficult-to-classify samples with low prediction probabilities. The Adam optimizer was initially set with a learning rate of 0.005, a momentum parameter of 0.9, and a second-order momentum parameter of 0.999. During training, the network adjustment coefficients are calculated every 80 iterations. The overall accuracy of the validation set is divided by the baseline accuracy of 0.75 to obtain the normalized accuracy value. The standard deviation of the recall for each class is multiplied by the weight coefficient of 2.5 and the reciprocal is taken to obtain the recall balance value. The rate of decrease of the loss function is divided by the baseline rate of decrease to obtain the normalized convergence rate value. The network adjustment coefficients are obtained by summing these values ​​according to the weight ratios of 0.5, 0.3, and 0.2. Training stops when the accuracy of the validation subset no longer improves for 10 consecutive iterations.

[0104] Technicians constructed a four-ary classification model for healthy leaves and leaves infected with three viruses: CMV, PVY, and TMV, and trained it using five cross-validation methods. Table 2 shows the confusion matrix classification results for different algorithm combinations, with the accuracy of all models ranging from 51% to 71%. Table 3 lists the detailed evaluation metrics for the top three accurate models.

[0105] Table 2. Performance scores of 16 model combinations in four categories (CK, CMV, PVY, TMV):

[0106]

[0107] Table 3 Evaluation results of quadruple classification models based on different feature extraction methods

[0108]

[0109] The SPA+RF model performed best in health category recognition, achieving an F1 score of 0.854, but its ability to recognize the TMV category was weak, with an F1 score of only 0.645. The GA+BRT model performed best in PVY category recognition, with an F1 score of 0.724, showing balanced overall performance. The CARS+RF model achieved an F1 score of 0.844 in the health category, but its low recall rate for CMV posed a risk of missed detections.

[0110] Technicians further developed an octet classification model for disease severity, categorizing healthy leaves and leaves infected with the three viruses into severity levels from 0 to 7. Table 4 lists the evaluation metrics of the top three models in terms of accuracy.

[0111] Table 4 Evaluation results of octal classification models based on different feature extraction methods

[0112]

[0113] The WOA+RF model exhibits the best overall performance, achieving an accuracy of 0.860 and the highest average F1 score. Its F1 score exceeds 0.85 across multiple disease stages, demonstrating high classification stability. The model achieves over 90% accuracy in classifying extreme disease levels 0 and 7, while maintaining a discrimination capability of over 75% for intermediate levels.

[0114] Technicians collected five groups of tobacco leaves for virus type and severity level identification tests. Following established procedures, they collected hyperspectral reflectance data, thermal infrared temperature field data, and chlorophyll fluorescence dynamics curve data. After Savitzky-Golay filtering, fractional-order differential enhancement, topological manifold embedding, and multimodal attention fusion processing, the resulting fused feature vectors were obtained. These fused feature vectors were then input into a trained WOA+RF virus classification and identification model. The model output showed that the first group consisted of healthy leaves, the second group had a CMV virus infection level of 3, the third group had a PVY virus infection level of 5, the fourth group had a TMV virus infection level of 2, and the fifth group had a PVY virus infection level of 7, completely consistent with the actual virus inoculation records and disease development stages.

[0115] This invention represents a significant technological advancement over traditional virus identification methods based on single spectral features. Traditional methods rely solely on point-by-point numerical changes in the spectral reflectance curve for feature extraction, making them susceptible to interference from environmental factors such as light variations and leaf angles. Furthermore, they lack sufficient sensitivity to detect subtle spectral differences in the early stages of disease. This invention, through fractional-order differential spectral enhancement processing, continuously adjusts the filtering characteristics of the differential operator in the frequency domain, achieving adaptive feature extraction for different disease types and severity levels. This significantly amplifies the response of early-stage chlorophyll content changes and cell structure damage on the spectral curve. The spectral topological manifold embedding process overcomes the limitations of traditional linear dimensionality reduction methods. By capturing the global morphological features and multi-scale geometric structure of the spectral curve through persistent cohomology eigenvalues ​​and Betti number sequence values, the topological invariants depend only on the relative geometric relationships of the spectral curve rather than absolute values, exhibiting good robustness to measurement noise. The multimodal attention fusion module integrates three complementary information sources: hyperspectral reflectance, thermal infrared temperature field, and chlorophyll fluorescence dynamics. Through an attention weight matrix, it adaptively allocates the feature contribution of different modes, comprehensively characterizing the biochemical metabolic disorders caused by viral infection. The dynamic network adjustment mechanism adjusts the learning rate and network layer number parameters in real time based on the performance of the validation set, avoiding the problem of insufficient generalization ability of fixed network structures on samples of different disease types and severity, and significantly improving the model's recognition accuracy for difficult-to-classify samples and the discrimination stability of different virus types.

[0116] It should be noted that the variables involved in this invention are explained in detail in Tables 5 and 6.

[0117] Table 5. Variable Explanation Table (Part 1)

[0118]

[0119] Table 6. Variable Explanation Table (Part Two)

[0120]

[0121] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a tobacco virus classification model based on tobacco leaf hyperspectral imaging, characterized in that, The process includes collecting hyperspectral reflectance data of tobacco leaves in the wavelength range of 380nm to 1000nm, performing Savitzky-Golay filtering to denoise the data, obtaining denoised hyperspectral reflectance data, performing fractional-derivative spectral enhancement on the denoised hyperspectral reflectance data to obtain fractional-derivative enhanced spectral data, performing spectral topological manifold embedding on the fractional-derivative enhanced spectral data to calculate persistent homology eigenvalues ​​and Betti number sequence values, and obtaining topological embedding feature vectors through dimensionality reduction using kernel principal component analysis, collecting thermal infrared temperature field data and chlorophyll fluorescence dynamics curve data corresponding to tobacco leaves, inputting the topological embedding feature vectors and thermal infrared temperature field data and chlorophyll fluorescence dynamics curve data into a multimodal attention fusion module to obtain fused feature vectors, counting the number of samples in each class, and using a spectral generative adversarial network to generate synthetic fused feature vectors for minority class samples, and using the fused feature vectors and synthetic fused feature vectors to form a training dataset to input into a virus classification and recognition model for training, using a focal loss function to calculate the classification loss value during training, and dynamically adjusting the learning rate parameter and network layer number parameter according to the network adjustment coefficient.

2. The method according to claim 1, characterized in that, Fractional differential spectral enhancement processing adaptively enhances the spectral reflectance curve based on fractional calculus theory. The order α of the fractional derivative is a real number between 0 and 1, and the Grünwald-Letnikov definition is used to calculate the α-order fractional derivative at the wavelength.

3. The method according to claim 2, characterized in that, The fractional order α is selected by adaptively searching on the training set using the particle swarm optimization algorithm. The objective function is defined as the ratio of the standard deviation of the inter-class spectral distance to the standard deviation of the intra-class spectral distance. Each particle in the particle swarm represents a candidate fractional order.

4. The method according to claim 3, characterized in that, The spectral topological manifold embedding process maps the global geometric structure of high-dimensional spectral data to a low-dimensional feature space based on algebraic topology theory. It treats the spectral reflectance curves of 480 bands as a set of points in a 480-dimensional Euclidean space and constructs the Vietoris-Rips complex of the point set.

5. The method according to claim 4, characterized in that, By gradually increasing the neighborhood radius parameter, a simplex structure with side triangles and tetrahedrons is generated from the discrete point set. The Betti number sequence values ​​of the complex under different neighborhood radii are calculated. The zeroth dimension component of the Betti number sequence value represents the number of connected components of the spectral data, and the first dimension component describes the ring topological structure features in the spectral curve.

6. The method according to claim 5, characterized in that, A persistent graph is plotted to record the generation and decay of topological features. Long-lived topological feature points are extracted from the persistent graph to construct topological descriptor vectors. Kernel principal component analysis is used to embed the topological descriptor vectors from the high-dimensional topological space into the low-dimensional Euclidean space. The radial basis function is selected as the kernel function.

7. The method according to claim 6, characterized in that, The multimodal attention fusion module includes a spectral feature encoding submodule, a thermal infrared feature encoding submodule, and a fluorescence feature encoding submodule. The spectral feature encoding submodule maps the topological embedding feature vector into a spectral intermediate representation vector through a fully connected layer.

8. The method according to claim 7, characterized in that, The thermal infrared feature encoding submodule extracts temperature gradient features from thermal infrared temperature field data through a convolutional neural network and maps them into thermal infrared intermediate representation vectors. The fluorescence feature encoding submodule extracts temporal decay features from chlorophyll fluorescence dynamics curve data through a long short-term memory network and maps them into fluorescence intermediate representation vectors.

9. The method according to claim 8, characterized in that, The attention weight matrix between the spectral intermediate representation vector, the thermal infrared intermediate representation vector, and the fluorescence intermediate representation vector is calculated. The attention weight matrix is ​​obtained through a scaled dot product attention mechanism. The weighted intermediate representation vector is obtained by weighting and summing the corresponding row vectors of each intermediate representation vector with the attention weight matrix.

10. The method according to claim 9, characterized in that, The virus classification and identification model uses a focus loss function to calculate the classification loss value. In the focus loss function, the modulation factor is defined as a power function of the predicted probability, and the power exponent ranges from 2.0 to 3.0.

Images