A method for monitoring severity of wheat scab based on UFD and triplet attention-1D-CNN

Abstract

The present application belongs to the technical field of crop disease monitoring, and particularly relates to a hyperspectral wheat scab severity monitoring method based on UFD and Triplet Attention-1D-CNN, comprising the following steps: S101. Adopting unwinding Fourier decomposition (UFD) to perform spectral enhancement processing on wheat canopy hyperspectral reflectance data; S102. Constructing an oversampling strategy based on class proportion IR regulation to perform balanced processing on the spectral samples; S103. Training a one-dimensional convolutional neural network model based on the balanced spectral samples to realize wheat scab severity grading identification. A complete technical scheme from data preprocessing, sample balancing to intelligent identification is formed, rapid, accurate and automatic grading monitoring of wheat scab severity is realized, and effective technical support is provided for precision agriculture decision-making.

Description

Technical Field

[0001] This invention belongs to the field of crop disease monitoring technology, specifically involving a hyperspectral method for monitoring the severity of wheat scab based on UFD and TripletAttention-1D-CNN. Background Technology

[0002] With the rapid development and application of UAV remote sensing, hyperspectral imaging technology, and deep learning algorithms in agriculture, agricultural production activities are gradually transforming towards mechanization, intelligence, and precision. Crop disease monitoring methods based on UAV hyperspectral data have become a research hotspot in this field, and have also provided a technical path to solve the shortcomings of traditional manual monitoring. In existing technologies, vegetation index analysis, principal component analysis (PCA), and continuous projection algorithm (SPA) are commonly used to extract features from hyperspectral data, and combined with models such as support vector machine (SVM), random forest (RF), or convolutional neural network (CNN) to achieve disease identification or classification. In the field of deep learning, one-dimensional convolutional neural network (1D-CNN) is widely used in hyperspectral classification tasks because it can directly perform end-to-end modeling of spectral sequences without complex manual feature engineering, and has become the mainstream deep learning model for monitoring the severity of wheat scab.

[0003] However, existing methods for monitoring the severity of wheat scab based on hyperspectral data still have certain shortcomings in practical applications:

[0004] (1) Insufficient utilization efficiency of hyperspectral signals. Existing methods mostly directly input the original spectral data or only perform simple smoothing and dimensionality reduction on the spectrum, failing to effectively mine the potential multi-scale structural information and disease-sensitive frequency domain features in the spectral signal, thus limiting the disease discrimination ability of the model;

[0005] (2) The problem of sample class imbalance is prominent. The severity samples of wheat scab under natural field conditions show obvious uneven distribution characteristics. The number of samples with high severity disease is significantly less than that of healthy or mild disease samples. The existing model is easily dominated by the majority class samples during the training process, resulting in low recognition accuracy and recall rate for high severity disease.

[0006] (3) The model has limited ability to focus on key spectral features. In the process of convolutional feature extraction, traditional 1D-CNN has relatively fixed weights for spectral features of different bands, making it difficult to adaptively highlight key spectral segments that are closely related to the severity of wheat scab, and thus failing to fully explore the disease discrimination value of full-band hyperspectral data.

[0007] To address the aforementioned issues, there is an urgent need to propose a targeted improvement solution to overcome the technical bottlenecks of existing methods.

[0008] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0009] The purpose of this invention is to provide a hyperspectral wheat scab severity monitoring method based on UFD and Triplet Attention-1D-CNN, so as to solve the problems existing in the prior art.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] A hyperspectral wheat scab severity monitoring method based on UFD and Triplet Attention-1D-CNN includes the following steps:

[0012] S101. Unwinding Fourier Decomposition (UFD) is used to perform spectral enhancement processing on the wheat canopy hyperspectral reflectance data acquired by the UAV. Enhanced spectral signal samples are obtained through frequency domain decomposition and finite term reconstruction.

[0013] S102. Construct an oversampling strategy based on category ratio IR control, and perform equalization processing on the enhanced spectral signal samples to obtain equalized spectral signal samples;

[0014] S103. Construct a one-dimensional convolutional neural network model with an embedded Triplet Attention mechanism, and train the one-dimensional convolutional neural network model based on the equalized spectral samples to achieve graded identification of wheat scab severity.

[0015] Furthermore, the hyperspectral wheat scab severity monitoring method using UFD and Triplet Attention-1D-CNN is characterized in that, in step S101, the Fourier decomposition includes processing the original wheat canopy hyperspectral reflectance curve. Perform frequency domain decomposition in complex exponential form;

[0016] ;

[0017] in, For the first The amplitude of each frequency component, For phase function, The order of the decomposition is... Let j be the imaginary unit, satisfying j 2 =-1;

[0018] The frequency components are filtered for principal energy components according to their energy proportions, and the filtered components are then further processed. The dominant frequency components are reconstructed to obtain the enhanced spectral signal. .

[0019] Furthermore, the main energy component is decomposed by setting a preset decomposition order. Truncation is performed to highlight disease-sensitive band information while maintaining the overall spectral structure; the decomposition order... satisfy:

[0020] ;

[0021] in, Set a preset energy retention threshold.

[0022] Furthermore, in step S102, the oversampling strategy includes counting the number of samples for each severity category. Calculate the class imbalance rate Set a target category ratio to make the number of samples in each category approach 1:1; set different oversampling factors for minority class samples to achieve a balanced distribution of categories;

[0023] The category imbalance rate The calculation formula is:

[0024] ;

[0025] in, This represents the number of samples in the category with the largest sample size.

[0026] Furthermore, the oversampling process employs the SMOTE method to construct synthetic spectral samples; the SMOTE method selects minority class samples and... A new sample is generated by linear interpolation between its nearest neighbors, represented as:

[0027] ,

[0028] in, For minority class samples, Its nearest neighbor samples .

[0029] Furthermore, in step S103, the Triplet Attention mechanism is embedded in the one-dimensional convolutional feature extraction process to adaptively weight and enhance the spectral features.

[0030] Furthermore, the one-dimensional convolutional neural network model includes three one-dimensional convolutional layers, each followed by a batch normalization layer and an activation function layer, and finally outputs the severity level of wheat scab through a fully connected layer and a Softmax function; wherein, the convolutional layer is used to extract local spectral features and to perform feature dimensionality reduction through a pooling layer.

[0031] Furthermore, the Triplet Attention mechanism includes a channel attention branch and a length attention branch; wherein, the attention branch generates feature representations in different directions by performing a dimension permutation operation on the input feature map; each branch generates attention weights by sharing a convolutional layer, and after normalization by the Sigmoid function, they are fused with the original feature map element by element.

[0032] This invention also provides a system applicable to the hyperspectral wheat Fusarium head blight severity monitoring method based on UFD and Triplet Attention-1D-CNN, including an input preprocessing module, a spectral decomposition and reconstruction module, a sample imbalance control module, an attention-enhanced one-dimensional convolutional feature extraction module, and a result output module. In the data flow, the acquired hyperspectral wheat canopy spectral data is first input to the preprocessing module for normalization and outlier filtering of the original spectra. Subsequently, the preprocessed spectral sequence enters the spectral decomposition and reconstruction module, which uses Unwinding Fourier Decomposition to perform multi-level decomposition of the spectral signal and reconstruct the disease-sensitive spectral features. The reconstructed spectral samples are input to the sample imbalance control module, which adjusts the distribution of samples at different severity levels. The adjusted samples are further input into a one-dimensional convolutional neural network embedded with a Triplet Attention module to complete feature extraction and disease severity discrimination. Finally, the output module provides the wheat Fusarium head blight severity classification result.

[0033] Among them, (1) the input preprocessing is as follows:

[0034] The collected hyperspectral data were processed on a per-ROI basis, and the average reflectance of each ROI in different spectral bands was extracted. The formula for the reflectance value is as follows:

[0035] In the formula, Indicates the reflectivity of the standard calibration cloth. Indicates canopy radiance. Indicates the emissivity of the standard calibration cloth. This represents the calculated reflectivity. A high-noise band is then proposed, resulting in a dataset.

[0036] (2) The implementation of the UFD module is as follows:

[0037] UFD is a signal decomposition method, given a real-valued periodic signal. Its Fourier series expansion is as follows

[0038]

[0039] make , yes Hardy spatial projection, also known as... in signal processing for The analyzed signal. Because Real value, and The following relationships hold true:

[0040]

[0041] in, Indicates taking The real part, yes The 0th Fourier expansion coefficient. This can be obtained through Plemelj's theorem in complex analysis theory:

[0042]

[0043] Where H is the Hilbert transform. Then for real signals... The study can be transformed into analyzing its signals. The research. According to mathematical complex analysis theory, analytic signals have the following Nevalinna decomposition theorem:

[0044]

[0045] in, , They are The outer function and the inner function. The inner function can be further decomposed into... ,

[0046]

[0047] It is by The Blaschke product consisting of all the zeros, yes The non-zero point. It is a singular inner function; for simplicity, we set it to 1 here, that is... A famous result in digital signal processing is that... and The energy is the same. Fourier series expansion ratio The convergence speed is fast. B is called the loop factor, and it is decomposed by separating B. This means "unwinding".

[0048] make ,but At least with For zero point, Decomposition yields:

[0049]

[0050] in, yes The Blaschke product consisting of the zeros of the product. Perform similar The decomposition yields:

[0051]

[0052] After repeating N times, we get:

[0053]

[0054] here ; ; It is the remainder term after N decompositions. Furthermore, let... for The sum of the first N terms of the UFD is then:

[0055]

[0056] Figure 2 The specific process of UFD is described. In fact, the classic Fourier series expansion is... Each time, a factor z is extracted. UFD is a generalization of Fourier series expansion, which extracts the loop factor to the maximum extent possible each time. This maximizes the convergence speed. Previous studies have theoretically demonstrated the exponential decay of the UFD convergence rate and calculated that UFD has the same computational complexity as the Fast Fourier Transform (FFT).

[0057] Since hyperspectral data often contains noise, this study uses a UFD-based method to approximate the intrinsic spectral curves in order to eliminate noise and reconstruct the spectral curves. However, it does not require a complete decomposition of the data to obtain all its information. Instead, we aim to use a minimal number of parameters. , It approximates the original spectral curve with extremely small errors. Therefore, any spectral curve... The analytic signal derived in equation (9) This can be represented as a finite linear combination plus a noise term. Wherein, It is a noise function. As shown in Figure 2, the UFD flowchart, the final result is... The first N terms of the UFD and From formula (5), we know the analytical signal The Nevalinna decomposition consists of an outer function and an inner function. Mai et al. proposed a method that first finds... The finite number of zeros is used to separate the finite Blaschke function, and then the corresponding outer function is obtained. Sun et al. proposed a method that first uses the mechanical quadrature method (MQM) to calculate the outer function, and verified through experiments that it is more suitable for non-stationary signals. Combining the two methods, a hyperspectral fitting algorithm is further implemented.

[0058] (3) Implementation of the SMOTE module

[0059] First, a minority class sample 'a' is selected from the original data. Then, the Euclidean distance between 'a' and the remaining minority class samples in the feature space is calculated to determine the K nearest neighbors. Based on the determined K nearest neighbors, the SMOTE algorithm randomly selects sample 'b', and finally generates a new synthetic sample using formula (11).

[0060]

[0061] The SMOTE algorithm has historically been used to address the problem of imbalanced minority class sample sizes in datasets. In recent years, as the effectiveness of SMOTE's superior oversampling strategy has been validated, increasing research has explored its applicability in synthetic sample generation. In this study, to ensure a near-balanced ratio of synthetic sample sizes across all classes (i.e., IR ≈ 1:1), instead of uniformly setting the generation factor for all minority classes, a different generation factor n is set based on the sample size of each minority class. By controlling the synthesis factor n for each minority class individually, the final sample size across all classes tends to be consistent, thereby expanding the sample size while optimizing the class distribution structure.

[0062] (4) Implementation of Triplet Attention-1D-CNN module

[0063] A three-layer convolutional neural network (CNN) was designed based on a combination of CNN and Triplet Attention. The network consists of three convolutional (Conv) layers, each followed by batch normalization and ReLU activation. A Triplet Attention module is introduced to dynamically adjust the channel and temporal weights of the features. Max pooling (pooling size 2) is then used for downsampling. Finally, the features are flattened and fed into two fully connected (FC) layers for classification output. Furthermore, to enhance the model's ability to capture feature importance, a Triplet Attention module specifically designed for one-dimensional data is introduced after each convolution. This module includes two sub-modules: Channel Attention (CA) and Length Attention (LA). Channel Attention generates channel attention weights by performing max pooling and average pooling on each channel, combined with convolution operations. Length Attention generates length attention weights by pooling the temporal data in the temporal dimension. Finally, the outputs of CA and LA are combined with convolutional features through element-wise multiplication to obtain weighted features. The weighted results of the two branches are then averaged and fused to ultimately improve the model's ability to represent complex spectral signals. The Triplet Attention-1D-CNN model framework diagram is shown below. Figure 3 As shown.

[0064] In addition, the key calculation formulas are explained below:

[0065] (1) UFD decomposition and reconstruction formula for hyperspectral signals

[0066] To address the issues of high noise levels and weak local features in the original hyperspectral reflectance signal, this invention employs Unwinding Fourier Decomposition (UFD) to decompose and reconstruct the input spectral signal. Its mathematical expression is as follows:

[0067]

[0068] in: This represents the original one-dimensional hyperspectral signal; Indicates the first Sub-signal components obtained from layer UFD decomposition; The number of decomposition layers; This is the residual signal after decomposition.

[0069] By selectively reconstructing multiple sub-components, a denoised and discriminative reconstructed spectral signal is obtained. This serves as input for subsequent deep models, thereby enhancing the spectral differences between different severity levels of Fusarium head blight.

[0070] (2) Formula for Triplet Attention Feature Enhancement Mechanism

[0071] To overcome the shortcomings of traditional one-dimensional convolution in modeling feature importance, this invention introduces a Triplet Attention mechanism in the feature extraction stage, which weights and enhances features from multiple dimensions. The formula for calculating the attention weights can be expressed as:

[0072]

[0073] in, The input feature tensor; This indicates a feature dimension rearrangement or pooling operation; Indicates convolution transformation; Use the Sigmoid activation function; This is the generated attention weight matrix.

[0074] The final output features are obtained in the following way:

[0075]

[0076] in, This indicates element-wise multiplication.

[0077] By introducing Triplet Attention, the model can highlight key band features that are highly correlated with the severity of Fusarium head blight and suppress redundant information.

[0078] Compared with the prior art, the present invention has the following beneficial effects:

[0079] (1) This invention introduces a spectral decomposition and reconstruction mechanism based on Unwinding Fourier Decomposition. By performing multi-level decomposition on the hyperspectral sequence, different frequency domain features are separated, and the component signals are combined according to the reconstruction strategy. The reconstruction can highlight the spectral expression of the sensitive change features of the disease, thereby strengthening the effective information and suppressing redundant features in the model input stage, solving the technical problem of insufficient mining of potential spectral structural features by existing methods.

[0080] (2) This invention introduces SMOTE to minority class samples while keeping the original majority class sample count unchanged, optimizes the synthetic sample generation strategy, and improves the class bias problem in the model parameter update process by constructing a balanced sample set with an imbalance ratio close to 1:1 to participate in model training. This improvement enables the model to fully learn the characteristics of high-severity diseases under imbalanced data conditions, and improves the computational stability and recognition ability of the model for minority class samples.

[0081] (3) The present invention embeds multi-branch attention enhancement in a one-dimensional convolutional feature extraction network, and weights the features at the channel dimension, spectral dimension and global interaction level, guiding the network to adaptively focus on spectral features that are highly correlated with the severity of the disease during the feature extraction process, thereby improving the model’s ability to discriminate complex spectral information and solving the problem of insufficient feature attention ability of existing models.

[0082] In summary, this invention first reconstructs the frequency domain of the original hyperspectral data through unwinding Fourier decomposition, effectively filtering out noise and enhancing the features of disease-sensitive bands. Then, it proposes an oversampling strategy based on class-ratio IR regulation, fundamentally optimizing the class distribution structure of the dataset and significantly improving the model's ability to identify minority classes (high-risk diseases). Finally, it constructs a one-dimensional convolutional neural network embedded with a Triplet Attention mechanism, adaptively focusing on key channels and band information in the spectral sequence to achieve automatic learning and extraction of deep-level discriminative features across the entire spectrum. This forms a complete technical solution from data preprocessing and sample balancing to intelligent identification, enabling rapid, accurate, and automated graded monitoring of wheat scab severity, providing effective technical support for precision agriculture decision-making. Attached Figure Description

[0083] Figure 1 This is a flowchart illustrating the wheat scab severity monitoring method based on UFD and Triplet Attention-1D-CNN of the present invention.

[0084] Figure 2 This is a schematic diagram of the UFD process of the present invention;

[0085] Figure 3 This is a schematic diagram of the structure of the improved Triplet Attention-1DCNN model based on this invention;

[0086] Figure 4 The following are schematic diagrams of the fitting plots of UFD, B-spline, and Fourier in (a) ROI1; (b) ROI24; (c) ROI83; (d) ROI121; and (e) ROI131 when the number of fitting iterations or internal nodes is 30. Detailed Implementation

[0087] The technical solution of this invention patent will be clearly and completely described below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0088] This embodiment addresses the problems of noise interference, uneven sample distribution, and insufficient utilization of full-band information in hyperspectral monitoring of wheat scab. It proposes a comprehensive method combining spectral reconstruction, sample equalization, and a deep learning model. Here, using UAV hyperspectral data from a wheat experimental area in Zhengzhou, Henan Province as an example, the proposed wheat scab severity monitoring method based on UFD and TripletAttention-1D-CNN is illustrated.

[0089] See Figure 1-4 A method for monitoring the severity of wheat scab based on UFD and Triplet Attention-1D-CNN includes the following steps:

[0090] S101. Perform spectral enhancement processing on the wheat canopy hyperspectral reflectance data acquired by the UAV, and use unwinding Fourier decomposition to perform frequency domain decomposition and finite term reconstruction on the hyperspectral data to extract disease-sensitive frequency components.

[0091] In this application, the acquired raw hyperspectral data (400-1000 nm) is susceptible to sensor noise and environmental factors, especially with a significant decrease in the signal-to-noise ratio in the band after 900 nm. To highlight sensitive information about disease and suppress noise, spectral enhancement preprocessing is required. Specifically, firstly, the high-noise bands after 900 nm are removed, retaining the effective band data of 400-900 nm. Subsequently, the unwound Fourier decomposition method is used to fit and reconstruct each spectral curve.

[0092] Suppose the original spectral signal is a real-valued periodic signal, and its analytic signal... It can be obtained through Hilbert transform. Using the UFD algorithm, Decomposed into a finite sum of principal energy components and a remainder:

[0093] ;

[0094] in, The decomposition coefficients are... For the first The Blaschke product (circling factor) obtained by the step. for The remainder after the decomposition (can be considered as noise).

[0095] Reconstructed signal (previous) Item Parts and and through Converting back to a real signal yields the reconstructed spectral curve. This process prioritizes the extraction of the main energy components of the signal through a "dewinding" mechanism, which can enhance subtle spectral changes caused by the disease while maintaining the overall spectral shape.

[0096] To determine the optimal number of fits The UFD method was compared with Fourier basis and B-spline basis methods. Experimental results show that the root mean square error (RMSE) of all three methods decreases with increasing fitting iterations, but UFD has the fastest convergence speed and the lowest error at the same number of fitting iterations. When the number of fitting iterations reaches 30, the reconstruction RMSE of UFD drops to approximately 0.6 × 10⁻³, significantly better than Fourier (1.7 × 10⁻³) and B-spline (4.6 × 10⁻³). Therefore, the selected fitting iteration number... This serves as a standard parameter for subsequent modeling, thereby achieving the best balance between smoothing noise and preserving details.

[0097] Step S102. To address the problem of sample class imbalance, an oversampling strategy based on class ratio IR regulation is constructed to equalize the spectral samples.

[0098] Due to the impact of field control measures, the distribution of sample numbers for different severity levels of wheat scab is extremely uneven. In the original dataset, the number of samples for severity levels 1 to 5 are 23, 59, 38, 14, and 10, respectively, with severely affected samples (levels 4 and 5) being severely scarce. To address this issue, this application employs synthetic minority class oversampling technology and innovatively introduces a strategy based on imbalance ratio control.

[0099] Specifically, first, calculate the sample size for each level. With the largest number of samples in the category imbalance ratio To ensure that the number of samples in each category tends to be balanced after oversampling (i.e., the target IR approaches 1:1), a uniform oversampling factor is not used. Instead, an independent generation factor is set for each minority category. Using the majority class (level 2, 59 samples) as a baseline, calculate the composite multiples for other classes: Level 1 Level 3: 2x, Level 4: 4x, Level 5: 6x. The SMOTE algorithm multiplies the value of each minority class sample in the feature space. Randomly select its K nearest neighbor samples And generate new samples using the following formula:

[0100]

[0101] in, It is a random number within the interval (0,1). This strategy generates a class-balanced dataset, fundamentally optimizing the class distribution structure.

[0102] Step S103. Construct a one-dimensional convolutional neural network model with embedded Triplet Attention mechanism, and train the model based on the equalized spectral samples to achieve graded identification of wheat scab severity.

[0103] In this application, to fully utilize full-band hyperspectral information and automatically learn deep discriminative features, a one-dimensional convolutional neural network model with an embedded Triplet Attention mechanism is constructed. This one-dimensional convolutional neural network model takes the spectral sequence (dimension 1 × number of bands) after UFD reconstruction and SMOTE equalization as input.

[0104] Specifically, the one-dimensional convolutional neural network model mainly consists of three one-dimensional convolutional layers: the first convolutional layer uses 16 filters of size 3, the second layer uses 32, and the third layer uses 64. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function, and downsampling is performed using a max pooling layer of size 2. The key improvement lies in embedding a Triplet Attention module after each convolutional operation. This module contains two parallel branches: a channel attention branch and a length attention branch. The channel attention branch performs max pooling and average pooling on the input feature map along the channel dimension, and then generates channel attention weights through a shared convolutional layer; the length attention branch performs similar pooling and convolution operations along the length dimension (i.e., band dimension) of the spectral sequence to generate length attention weights. The attention weights generated by the two branches are normalized by the Sigmoid function, multiplied element-wise with the original feature map, and finally the two weighted feature maps are averaged and fused. This mechanism enables the model to adaptively focus on key features in different channels and at different spectral positions.

[0105] At the end of the model, the 3D feature map is flattened into a 1D vector and input into a fully connected layer with 128 neurons, ReLU activation, and a dropout rate of 0.5. Finally, it is classified through a Softmax output layer with 5 neurons (corresponding to 5 disease levels). The model uses the cross-entropy loss function and is trained for 200 epochs.

[0106] The UFD-SMOTE-Triplet Attention-1D-CNN model (USTA-CNN) proposed in this application is compared with a combination model of various traditional feature selection methods (mutual information MI, discriminant coefficient DC, AFD) and multiple classifiers (GBDT, RF, XGBoost, KNN). The detection results are shown in Table 1.

[0107] Table 1 Comparison of detection results from different models

[0108]

[0109] As shown in Table 1, the best traditional combined model (UFD-MI-SMOTE-GBDT) has an accuracy of 0.85, while the USTA-CNN model of this application has an accuracy of 0.9333. All indicators are significantly better than all traditional model combinations, indicating that the deep learning model constructed in this application can more fully mine full-band spectral information and reduce the dependence on manual feature engineering.

[0110] The foregoing description of specific exemplary embodiments of the invention is for illustrative and explanatory purposes. These descriptions are not intended to limit the invention to the precise forms disclosed, and it will be apparent that many changes and variations can be made in accordance with the foregoing teachings. The exemplary embodiments were chosen and described in order to explain the specific principles of the invention and its practical application, thereby enabling those skilled in the art to implement and utilize various different exemplary embodiments of the invention, as well as various different choices and variations. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims

1. A hyperspectral method for monitoring the severity of wheat scab based on UFD and Triplet Attention-1D-CNN, characterized in that, Includes the following steps: S101. The wheat canopy hyperspectral reflectance data acquired by the UAV was enhanced by using unwound Fourier decomposition UFD, and the enhanced spectral signal sample was obtained by frequency domain decomposition and finite term reconstruction. S102. Construct an oversampling strategy based on category ratio IR control, and perform equalization processing on the enhanced spectral signal samples to obtain equalized spectral signal samples; S103. Construct a one-dimensional convolutional neural network model with an embedded Triplet Attention mechanism, and train the one-dimensional convolutional neural network model based on the equalized spectral samples to achieve graded identification of wheat scab severity.

2. The hyperspectral wheat scab severity monitoring method based on UFD and Triplet Attention-1D-CNN according to claim 1, characterized in that, In step S101, the Fourier decomposition includes processing the original wheat canopy hyperspectral reflectance curve. Perform frequency domain decomposition in complex exponential form; ； in, For the first The amplitude of each frequency component, For phase function, The order of the decomposition is... Let j be the imaginary unit, satisfying j 2 =-1; The frequency components are filtered for principal energy components according to their energy proportions, and the filtered components are then further processed. The dominant frequency components are reconstructed to obtain the enhanced spectral signal. .

3. The hyperspectral wheat scab severity monitoring method based on UFD and Triplet Attention-1D-CNN according to claim 2, characterized in that, The main energy component is decomposed by setting a preset decomposition order. Truncation is performed to highlight disease-sensitive band information while maintaining the overall spectral structure; the decomposition order... satisfy: ； in, Set a preset energy retention threshold.

4. The hyperspectral wheat scab severity monitoring method based on UFD and Triplet Attention-1D-CNN according to claim 1, characterized in that, In step S102, the oversampling strategy includes counting the number of samples for each severity category. Calculate the class imbalance rate ; Set a target category ratio so that the number of samples in each category approaches 1:1; set different oversampling factors for minority class samples to achieve a balanced distribution of categories; The category imbalance rate The calculation formula is: ； in, This represents the number of samples in the category with the largest sample size.

5. The hyperspectral wheat scab severity monitoring method based on UFD and Triplet Attention-1D-CNN according to claim 4, characterized in that, The oversampling process employs the SMOTE method to construct synthetic spectral samples; the SMOTE method selects minority class samples and... A new sample is generated by linear interpolation between its nearest neighbors, represented as: ， in, For minority class samples, Its nearest neighbor samples .

6. The hyperspectral wheat scab severity monitoring method based on UFD and Triplet Attention-1D-CNN according to claim 1, characterized in that, In step S103, the Triplet Attention mechanism is embedded in the one-dimensional convolutional feature extraction process to adaptively weight and enhance spectral features.

7. The hyperspectral wheat scab severity monitoring method based on UFD and Triplet Attention-1D-CNN according to claim 6, characterized in that, The one-dimensional convolutional neural network model includes three one-dimensional convolutional layers. Each convolutional layer is followed by a batch normalization layer and an activation function layer, and finally outputs the severity level of wheat scab through a fully connected layer and a Softmax function. The convolutional layers are used to extract local spectral features, and the pooling layer is used for feature dimensionality reduction.

8. The hyperspectral wheat scab severity monitoring method based on UFD and Triplet Attention-1D-CNN according to claim 6, characterized in that, The Triplet Attention mechanism includes a channel attention branch and a length attention branch. The attention branch generates feature representations in different directions by performing a dimension permutation operation on the input equalized spectral signal samples. Each branch generates attention weights through a shared convolutional layer, and after normalization by the Sigmoid function, it is fused with the original feature map element by element.

9. A system applicable to the hyperspectral wheat scab severity monitoring method based on UFD and Triplet Attention-1D-CNN as described in any one of claims 1-8, characterized in that, include: The input preprocessing module normalizes and filters outliers from the raw spectra of the acquired hyperspectral wheat canopy spectral data. The spectral decomposition and reconstruction module uses unwound Fourier decomposition UFD to perform multi-level decomposition on the preprocessed spectral signal and reconstruct the disease-sensitive spectral features. A sample imbalance control module, which adjusts the distribution of samples at different severity levels; An attention-enhanced one-dimensional convolutional feature extraction module is provided, which performs feature extraction and disease severity determination. The result output module outputs the severity grading results of wheat scab. The result input module outputs the severity grading results of wheat scab.

Images