Multi-variety Citrus Huanglongbing detection method and system based on hyperspectral imaging

Abstract

The invention provides a multi-variety citrus Huanglongbing detection method and system based on hyperspectral imaging. The method comprises the following steps: collecting hyperspectral image samples containing different varieties of healthy and diseased citrus leaves; dividing a region of interest of the hyperspectral image sample, and exporting spectral data of the region of interest; according to the root-mean-square error of the average spectrum curve of the healthy hyperspectral image sample and the diseased hyperspectral image sample, preliminarily screening in the spectrum data to determine the number N of wave band groups; performing secondary screening on the N wave bands to determine M core characteristic wave bands; constructing a discrimination model by using a convolutional neural network, and training the discrimination model by using waveband data of M core characteristic wavebands in the healthy hyperspectral image sample and the diseased hyperspectral image sample; the trained discrimination model is used for predicting the detection result of the Candidatus Liberobacter asiaticum. According to the method, rapid field screening compatible with multiple varieties is achieved, and a standardized technical scheme is provided for digital plant protection of different citrus producing areas.

Description

Technical Field

[0001] The present invention belongs to the technical field of crop disease prevention and control, and relates to a method and system for detecting huanglongbing of multiple citrus varieties based on hyperspectral imaging. Background Art

[0002] Huanglongbing (HLB) is a devastating disease caused by phloem bacteria, and its early detection is crucial for prevention and control. At present, hyperspectral imaging technology (400 - 1000 nm) has been used for detecting diseases of single citrus varieties. By analyzing the differences in leaf spectral reflectance (for example, the reflectance of healthy leaves at the chlorophyll absorption valley of 550 nm is 12% - 15% higher than that of diseased leaves), combined with machine learning algorithms, a classification accuracy of more than 85% can be achieved within a single variety. However, there are serious technical bottlenecks in the multi - variety compatible detection of this technology. Existing hyperspectral detection technologies are limited by variety - specific spectral response mechanisms, feature selection biases, and data heterogeneity, and are essentially "single - variety closed systems". When directly extending the single - variety model to the multi - variety scenario, due to the lack of cross - variety spectral alignment mechanisms, adaptive feature screening frameworks, and heterogeneous data fusion strategies, the detection efficiency drops sharply. This theoretically proves that the technical route based on the traditional hyperspectral analysis framework cannot break through the technical barriers of multi - variety compatible detection.

[0003] Therefore, developing a hyperspectral citrus huanglongbing detection method that adapts to multi - variety differences and solving the bottleneck of the insufficient generalization ability of existing models are urgent problems to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, the present invention proposes a method and system for detecting huanglongbing of multiple citrus varieties based on hyperspectral imaging. Through a spectral band adaptive screening algorithm, it breaks through the technical bottleneck of the insufficient generalization ability of the single - variety detection model, realizes rapid field screening compatible with multiple varieties, and provides a standardized technical solution for digital plant protection in different citrus production areas.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] The present invention first discloses a method for detecting huanglongbing of multiple citrus varieties based on hyperspectral imaging, including the following steps:

[0007] Model training step:

[0008] S1: Collect hyperspectral image samples of healthy and diseased citrus leaves containing different varieties;

[0009] S2: Divide the region of interest of the hyperspectral image samples and export the spectral data of the region of interest;

[0010] S3: Based on the root mean square error between the average spectral curves of healthy hyperspectral image samples and diseased hyperspectral image samples, preliminarily screen and determine the number N of band groups in the spectral data, where N > 2;

[0011] S4: Perform a secondary screening on the N bands to determine M core feature bands, where N > M ≥ 2;

[0012] S5: Use a convolutional neural network to construct a discrimination model, and use the band data of the M core feature bands in the healthy hyperspectral image samples and diseased hyperspectral image samples as training samples to train the discrimination model;

[0013] Model detection steps:

[0014] S6: Input the hyperspectral image of the citrus leaf to be detected into the trained discrimination model, and output the detection result of citrus huanglongbing.

[0015] Preferably, the step of dividing the region of interest of the hyperspectral image sample in S1 includes: avoiding the vein region in the hyperspectral image sample, and dividing the regions of interest in the upper, middle, and lower parts of the leaf.

[0016] Preferably, before S3, it also includes a preprocessing step for the spectral data:

[0017] Use a Savitzky-Golay filter to perform denoising on the spectral data, and then use a standard normal transformation to eliminate the scattering interference of the denoised spectral data to obtain the preprocessed spectral data for performing the S3 operation.

[0018] Preferably, S3 includes:

[0019] S31: Based on the entire spectral range of the hyperspectral image sample, select P groups of feature band combinations, where the feature band values within each of the P groups of feature band combinations are different, and the number of feature bands between each combination is also different, and P > N;

[0020] S32: For each feature band combination, calculate its average spectral curve, and compare the root mean square error of the corresponding feature band combination in the healthy samples and non-healthy samples;

[0021] S33: Select the number of the feature band combination with the smallest root mean square error as the number N of the preliminarily screened and determined band groups.

[0022] Preferably, S4 includes:

[0023] S41: Use the SPA algorithm to screen N feature band values;

[0024] S42: Perform secondary screening among the N characteristic band values through the particle swarm optimization algorithm to determine M core characteristic bands.

[0025] Preferably, the convolutional neural network architecture sequentially includes, according to the data processing order:

[0026] An input layer, which reconstructs the spectral reflectance and spatial texture features of the M core characteristic bands into a two-channel time series as the input, uses a one-dimensional convolutional kernel with 64 channels, combines the ReLU activation function and applies L2 regularization;

[0027] A max pooling layer, which is used to compress the feature dimension;

[0028] A flattening layer, which is used to realize the conversion from spatial features to vectors;

[0029] A fully connected layer, which adopts the Dropout mechanism;

[0030] An output layer, which outputs the disease probability value through the Sigmoid activation function.

[0031] Preferably, the training process of the discriminant model uses the Adam optimizer, takes minimizing the binary cross-entropy loss as the loss function, and implements an early stopping strategy during the training process.

[0032] Preferably, in S41, the SPA algorithm is used to screen 7 characteristic band values, which are 68, 93, 134, 149, 155, 163, and 186, and the corresponding wavelengths are 536nm, 586nm, 668nm, 698nm, 710nm, 726nm, and 772nm respectively.

[0033] Preferably, in S42, the two core characteristic bands determined by secondary screening are the 155th band and the 163rd band, and the corresponding wavelengths are 710nm and 726nm respectively.

[0034] The present invention also discloses a multi-variety citrus huanglongbing detection system according to the above-mentioned multi-variety citrus huanglongbing detection method based on hyperspectral imaging, including:

[0035] A hyperspectral acquisition module, which is used to acquire hyperspectral image samples of healthy and diseased citrus leaves of different varieties;

[0036] An ROI division module, which is used to divide the region of interest of the hyperspectral image sample and export the spectral data of the region of interest;

[0037] A characteristic band confirmation module, which is used to preliminarily screen and determine the number N of band groups in the spectral data according to the root mean square error of the average spectral curves of the healthy hyperspectral image sample and the diseased hyperspectral image sample, where N>2; perform secondary screening on the N bands to determine M core characteristic bands, where N>M≥2;

[0038] A discriminant model construction module, which is used to construct a discriminant model using a convolutional neural network, and uses the band data of the M core feature bands in the healthy hyperspectral image samples and the diseased hyperspectral image samples as training samples to train the discriminant model;

[0039] A citrus huanglongbing detection module, which is used to input the hyperspectral image of the citrus leaf to be detected into the trained discriminant model and output the citrus huanglongbing detection result.

[0040] As can be seen from the above technical solutions, compared with the prior art, the beneficial effects of the present invention include:

[0041] By using the Savitzky-Golay filter and standard normal variate (SNV) for joint denoising and standardization processing, the present invention can effectively reduce the noise in the spectral data and the interference of environmental factors. This improves the accuracy and reliability of subsequent analysis.

[0042] The present invention uses the particle swarm optimization algorithm (PSO) to optimize and screen the feature bands, and finally determines the core feature bands. This method reduces the dimension of the input data, simplifies the model complexity, thereby improving the calculation efficiency and the model running speed, and reducing the resource consumption.

[0043] The dual-channel spectral-spatial feature fusion model constructed by the present invention based on the convolutional neural network (CNN) has a classification accuracy significantly better than that of traditional methods such as the support vector machine (SVM) and the BP neural network. A higher accuracy means more effective disease prevention and control measures, which can timely detect and handle diseased citrus trees, avoid the spread of diseases, and reduce the economic losses caused by diseases in the long run.

[0044] The method proposed by the present invention is applicable to multiple main cultivated citrus varieties and realizes cross-variety compatible detection. This feature greatly expands the application scope of the technology, reduces the need to develop independent detection schemes for different varieties, and further saves the R & D and deployment costs.

[0045] In summary, the present invention not only improves the accuracy and efficiency of citrus huanglongbing detection, but also shows significant advantages in reducing costs. This is of great significance for the development of precision agriculture. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts;

[0047] Figure 1 This is a flowchart of the method for detecting huanglongbing of multiple citrus varieties based on hyperspectral imaging provided by the embodiments of the present invention;

[0048] Figure 2 This is a schematic diagram of the specific execution process of the method for detecting huanglongbing of multiple citrus varieties based on hyperspectral imaging provided by the embodiments of the present invention;

[0049] Figure 3 This is a schematic diagram of the ROI area selected for healthy ponkan leaves provided by the embodiments of the present invention;

[0050] Figure 4 This is a schematic diagram of various spectral curves of healthy ponkan leaves in ENVI 5.6 software provided by the embodiments of the present invention;

[0051] Figure 5 This is a schematic diagram of the original spectral curves of healthy samples and non-healthy samples of five citrus varieties provided by the embodiments of the present invention;

[0052] Figure 6 This is a schematic diagram of the comparison of the average spectral curves of healthy samples and non-healthy samples of five citrus varieties after SG+SNV preprocessing provided by the embodiments of the present invention;

[0053] Figure 7 This is a schematic diagram of the optimal result of the number of selected characteristic bands calculated by the root mean square error provided by the embodiments of the present invention;

[0054] Figure 8 This is a schematic diagram of the characteristic bands selected by the SPA algorithm provided by the embodiments of the present invention.

[0055] Figure 9 This is a schematic diagram of the characteristic bands optimized by the PSO algorithm provided by the embodiments of the present invention.

[0056] Figure 10 This is a schematic diagram of the confusion matrix and ROC curve of the CNN model provided by the embodiments of the present invention.

[0057] Figure 11 This is a schematic diagram of a three-dimensional bar chart of the multi-index performance comparison of SVM, BP neural network and CNN algorithms provided by the embodiments of the present invention. Detailed implementation manners

[0058] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0059] In the first aspect of the embodiment of the present invention, a detection method for Huanglong disease of multiple varieties of citrus based on hyperspectral imaging is provided. By constructing a three-dimensional spectral database containing five main cultivated varieties, the ROI three-region sampling method is used to eliminate spatial heterogeneity; a two-stage spectral correction module is designed; an SPA-PSO-RMSE adaptive feature selection system is developed to screen out a feature wavelength subset compatible with the variety; and a two-channel CNN model with a variety attention mechanism is created. As Figure 1 shown, it includes the following steps:

[0060] Model training steps:

[0061] S1: Collect hyperspectral image samples of healthy and diseased citrus leaves containing different varieties; for example, collect healthy and diseased citrus leaf samples of five main cultivated citrus varieties including Benidah 1, Benidah 2, Wogan, Grapefruit, and Ponkan, and use a hyperspectral imaging system (spectral range 400 - 1000 nm) to obtain the hyperspectral image data of the leaves.

[0062] S2: Divide the region of interest of the hyperspectral image samples and export the spectral data of the region of interest;

[0063] S3: According to the root mean square error of the average spectral curves of the healthy hyperspectral image samples and the diseased hyperspectral image samples, preliminarily screen and determine the number N of band groups in the spectral data, N > 2;

[0064] S4: Perform a secondary screening on the N bands to determine M core feature bands, N > M ≥ 2;

[0065] S5: Use a convolutional neural network to construct a discriminant model, and use the band data of the M core feature bands in the healthy hyperspectral image samples and the diseased hyperspectral image samples as training samples to train the discriminant model;

[0066] Model detection steps:

[0067] S6: Input the hyperspectral image of the citrus leaf to be detected into the trained discriminant model, and output the detection result of Huanglong disease of citrus.

[0068] In one embodiment, the step of dividing the region of interest of the hyperspectral image samples in S1 includes: avoiding the vein region in the hyperspectral image samples, dividing the regions of interest in the upper, middle, and lower regions of the leaf. The ROI regions can be drawn in the upper, middle, and lower regions of the leaf using ENVI 5.6 software. As Figure 3 shown, export the spectral data in CSV format, as Figure 4 shown.

[0069] It should be noted that the above spectral data export operations are respectively performed on the healthy leaves and diseased leaves of five varieties, as Figure 5 shown, obtaining 10 CSV files.

[0070] In one embodiment, as Figure 2 shown, before S3, there is also a preprocessing step for the spectral data:

[0071] Use the Savitzky-Golay filter to denoise the spectral data, and then use the standard normal transformation to eliminate the scattering interference of the denoised spectral data, amplify the feature differences, and make the reflectivity contrast of the key bands more significant. Obtain the preprocessed spectral data for performing the S3 operation.

[0072] In one embodiment, S3 includes:

[0073] S31: Based on the entire spectral range of the hyperspectral image samples, select P groups of characteristic band combinations, where the characteristic band values within each of the P groups of characteristic band combinations are different, and the number of characteristic bands between each combination is different, P > N;

[0074] S32: For each characteristic band combination, respectively merge its corresponding healthy samples and non-healthy samples, calculate the average spectral curves of the healthy samples and non-healthy samples, as Figure 6 shown, and compare the root mean square error of the corresponding characteristic band combinations in the healthy samples and non-healthy samples, as Figure 7 shown;

[0075] S33: Select the number of characteristic band combinations with the smallest root mean square error as the initially screened and determined number of band groups N.

[0076] In one embodiment, as Figure 2 shown, S4 includes:

[0077] S41: Use the SPA algorithm to screen N characteristic band values, as Figure 8 shown;

[0078] S42: Through the particle swarm optimization algorithm PSO, perform a secondary screening among the N characteristic band values to determine M core characteristic bands, as Figure 9 shown.

[0079] In this embodiment, when the calculation result in S41 is 7, the corresponding RMSE is the smallest and there is no significant change subsequently. Using the SPA algorithm to screen 7 characteristic band values are 68, 93, 134, 149, 155, 163 and 186, and the corresponding wavelengths are 536nm, 586nm, 668nm, 698nm, 710nm, 726nm and 772nm respectively.

[0080] In this embodiment, in S42, the results of SPA feature selection are globally optimized to screen out a better band combination, and finally optimized to two core feature bands through the PSO algorithm, namely the 155th band and the 163rd band, corresponding to wavelengths of 710 nm and 726 nm respectively.

[0081] In one embodiment, the convolutional neural network architecture sequentially includes, according to the data processing order:

[0082] An input layer that reconstructs the spectral reflectance and spatial texture features of M core feature bands into a two-channel time series as a two-dimensional tensor (number of samples × 2 × 1) for input, uses a one-dimensional convolutional kernel with 64 channels (kernel_size = 2), combines a ReLU activation function and applies L2 regularization (λ = 0.01); the convolutional kernel constrained by L2 regularization extracts local correlation features.

[0083] A max-pooling layer for compressing the feature dimension.

[0084] A flattening layer for converting spatial features into vectors, with a stride of 1.

[0085] A fully connected layer with 32 neurons (L2 regularization λ = 0.005), combined with a 40% Dropout mechanism to prevent overfitting.

[0086] An output layer that outputs the disease probability value through a Sigmoid activation function.

[0087] In one embodiment, the training process of the discriminant model uses an Adam optimizer with a learning rate of 0.0002, uses minimizing binary cross-entropy loss as the loss function, and implements an early stopping strategy during the training process (patience value of 15 rounds, minimum improvement threshold of 0.001).

[0088] In this embodiment, when the output probability exceeds the 0.5 decision threshold, the diagnostic sample is positive for huanglongbing, and the area under the ROC curve (AUC) corresponding to this threshold is used as the core index of the model's discriminative power, as Figure 10 shown.

[0089] In the second aspect of the embodiments of the present invention, a multi-variety citrus huanglongbing detection system for a multi-variety citrus huanglongbing detection method based on hyperspectral imaging according to the first aspect of the embodiments is provided, including:

[0090] A hyperspectral acquisition module for acquiring hyperspectral image samples of healthy and diseased citrus leaves of different varieties;

[0091] An ROI division module for dividing the region of interest of the hyperspectral image samples and exporting the spectral data of the region of interest;

[0092] A characteristic band confirmation module, which is used to preliminarily screen and determine the number of band groups N in the spectral data according to the root mean square error between the average spectral curves of healthy hyperspectral image samples and diseased hyperspectral image samples, where N > 2; and perform a secondary screening on the N bands to determine M core characteristic bands, where N > M ≥ 2;

[0093] A discriminant model construction module, which is used to construct a discriminant model using a convolutional neural network, and use the band data of the M core characteristic bands in the healthy hyperspectral image samples and diseased hyperspectral image samples as training samples to train the discriminant model;

[0094] A citrus huanglongbing detection module, which is used to input the hyperspectral image of the citrus leaf to be detected into the trained discriminant model and output the citrus huanglongbing detection result.

[0095] In the CNN model for 150 test set samples, among 75 healthy samples, 74 were correctly identified, and one was misjudged as diseased; among 75 diseased samples, 68 were correctly identified, and 7 were missed and detected as healthy samples. 142 samples were correctly identified, with an accuracy rate of 94.67%, a recall rate of 90.67%, a precision rate of 98.55%, an F1 value of 94.44%, and an AUC of 99.54%. The model's ability to distinguish between the two types of samples is nearly perfect, and the prediction confidence is reliable. As Figure 11 shown, compared with the SVM model and the BP neural network model, multi-dimensional verification from the accuracy rate (94.67%) to the F1 value (94.44%) and other aspects shows that the CNN model has comprehensive performance advantages compared with the traditional SVM and BP neural networks.

[0096] The above has introduced in detail the multi-variety citrus huanglongbing detection method and system based on hyperspectral imaging provided by the present invention. In this embodiment, specific examples are used to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

[0097] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined in this embodiment can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown in this embodiment, but will conform to the widest scope consistent with the principles and novel features disclosed in this embodiment.

Claims

1. A detection method for Huanglongbing of multiple citrus varieties based on hyperspectral imaging, characterized in that, It includes the following steps: Model training step: S1: Collect hyperspectral image samples of healthy and diseased citrus leaves of different varieties; S2: Divide the region of interest of the hyperspectral image samples and export the spectral data of the region of interest; S3: Based on the root mean square error of the average spectral curves of healthy hyperspectral image samples and diseased hyperspectral image samples, preliminarily screen and determine the number N of band groups in the spectral data, where N > 2; S4: Perform secondary screening on the N bands to determine M core feature bands, where N > M ≥ 2; S5: Use a convolutional neural network to construct a discriminant model, and use the band data of the M core feature bands in healthy hyperspectral image samples and diseased hyperspectral image samples as training samples to train the discriminant model; Model detection step: S6: Input the hyperspectral image of the citrus leaf to be detected into the trained discriminant model and output the detection result of citrus huanglongbing.

2. The multi-variety citrus huanglongbing detection method based on hyperspectral imaging according to claim 1, wherein, The step of dividing the region of interest of the hyperspectral image samples in S1 includes: avoiding the vein region in the hyperspectral image samples and dividing the regions of interest in the upper, middle, and lower regions of the leaf.

3. A method for detecting Huanglongbing of multiple varieties of citrus based on hyperspectral imaging according to claim 1, characterized in that, Before S3, there is also a preprocessing step for the spectral data: Use a Savitzky-Golay filter to denoise the spectral data, and then use a standard normal transformation to eliminate the scattering interference of the denoised spectral data to obtain the preprocessed spectral data for performing the operation of S3.

4. A method for detecting Huanglongbing of multiple citrus varieties based on hyperspectral imaging according to claim 1, characterized in that, S3 includes: S31: Based on the entire spectral range of the hyperspectral image samples, select P groups of feature band combinations, where the feature band values within each of the P groups of feature band combinations are different, and the number of feature bands between each combination is different, and P > N; S32: For each feature band combination, calculate its average spectral curve and compare the root mean square error of the corresponding feature band combination in healthy samples and non-healthy samples; S33: Select the number of feature band combinations with the smallest root mean square error as the preliminarily screened and determined number N of band groups.

5. A method for detecting Huanglongbing of multiple varieties of citrus based on hyperspectral imaging according to claim 1, characterized in that, S4 includes: S41: Use the SPA algorithm to screen N feature band values; S42: Perform secondary screening on the N feature band values through a particle swarm algorithm to determine M core feature bands.

6. A method for detecting Huanglongbing of multiple citrus varieties based on hyperspectral imaging according to claim 1, characterized in that, The convolutional neural network architecture sequentially includes, according to the data processing order: Input layer, reconstruct the spectral reflectance and spatial texture features of the M core feature bands into a two-channel time series as input, use a one-dimensional convolutional kernel with 64 channels, combine with a ReLU activation function and apply L2 regularization; Max pooling layer, used to compress the feature dimension; Flattening layer, used to realize the conversion from spatial features to vectors; Fully connected layer, using a Dropout mechanism; Output layer, output the disease probability value through a Sigmoid activation function.

7. A method for detecting huanglongbing of multiple varieties of citrus based on hyperspectral imaging according to claim 1, characterized in that The training process of the discriminant model uses an Adam optimizer, uses the minimization of binary cross-entropy loss as the loss function, and implements an early stopping strategy during the training process.

8. A method for detecting Huanglongbing of multiple citrus varieties based on hyperspectral imaging according to claim 5, characterized in that, In S41, the SPA algorithm is used to screen 7 characteristic band values of 68, 93, 134, 149, 155, 163, and 186, and the corresponding wavelengths are 536nm, 586nm, 668nm, 698nm, 710nm, 726nm, and 772nm respectively.

9. The method for detecting huanglongbing of multiple citrus varieties based on hyperspectral imaging according to claim 8, wherein, In S42, two core characteristic bands are determined by secondary screening as the 155th band and the 163rd band, and the corresponding wavelengths are 710nm and 726nm respectively.

10. A multi-variety citrus huanglongbing detection system for a multi-variety citrus huanglongbing detection method based on hyperspectral imaging according to any one of claims 1-9, characterized in that, Including: A hyperspectral acquisition module for acquiring hyperspectral image samples of healthy and diseased citrus leaves of different varieties; An ROI division module for dividing the region of interest of the hyperspectral image sample and exporting the spectral data of the region of interest; A characteristic band confirmation module for preliminarily screening and determining the number N of band groups in the spectral data according to the root mean square error of the average spectral curves of the healthy hyperspectral image sample and the diseased hyperspectral image sample, N > 2; performing secondary screening on the N bands to determine M core characteristic bands, N > M ≥ 2; A discriminant model construction module for constructing a discriminant model using a convolutional neural network, and using the band data of the M core characteristic bands in the healthy hyperspectral image sample and the diseased hyperspectral image sample as training samples to train the discriminant model; A citrus huanglongbing detection module for inputting the hyperspectral image of the citrus leaf to be detected into the trained discriminant model and outputting the detection result of citrus huanglongbing.

Images