A citrus huanglongbing detection method based on fusion of visual and olfactory information

Abstract

This invention provides a method for detecting Huanglongbing (HLB) in citrus based on the fusion of visual and olfactory information, belonging to the field of plant disease detection technology. It solves the technical problem that single-sensor technology is insufficient to accurately identify different symptom stages of HLB, and that early-stage HLB and zinc-deficient HLB samples are prone to missed detection and misjudgment. The technical solution includes the following steps: collecting citrus leaf samples and completing gas collection; acquiring reflectance and transmission polarization images using a computer vision device, and extracting gray-level co-occurrence matrix texture features and gray-level histogram features; collecting leaf gases using an electronic nose device and extracting extreme value features; using an improved recursive feature elimination and cross-validation (RFECV) method to fuse visual and olfactory features at a feature layer; establishing a stepwise classification model, and identifying each classification step based on the linear discriminant analysis (LDA) algorithm. This invention achieves accurate identification of samples at different symptom stages of HLB through the fusion of visual and olfactory multimodal features.

Description

Technical Field

[0001] This invention relates to the field of plant disease detection technology, and in particular to a method for detecting Huanglongbing (HLB) of citrus based on the fusion of visual and olfactory information. Background Technology

[0002] Citrus fruits are rich in nutrients and have a unique flavor. The citrus industry faces various pest and disease threats, among which Huanglongbing (HLB), a highly contagious and damaging disease, has been listed as a legally mandated quarantine disease. HLB is mainly transmitted through the Asian citrus psyllid and grafting. Psyllids can transmit the pathogen within 24 hours and spread it rapidly with the help of wind. HLB also has a long incubation period, allowing the disease to spread before symptoms appear. HLB has spread to more than 40 countries and regions worldwide, posing a serious threat to the citrus industry. Currently, there is no effective treatment; early detection and removal of diseased trees are the most effective control methods. Therefore, developing rapid and accurate HLB detection technology is of great significance for controlling the spread of the disease and protecting citrus resources.

[0003] Huanglongbing (HLB) symptoms are complex and difficult to describe precisely. Based on symptom stages, it can be divided into early, mid-stage (typical symptom stage), and late stage. Infected plants experience uneven blockage of the phloem, leading to starch accumulation in the photosynthetic tissue and sieve tubes of the leaves. Early-stage infected plants resemble healthy plants but already possess the ability to spread the disease. During the typical symptom stage, leaves exhibit mottled yellowing caused by large-area irregular starch accumulation. Late-stage infected plants suffer from root rot and decreased zinc absorption capacity, with leaf symptoms resembling zinc deficiency, easily masking the characteristics of HLB. Due to the complexity of the symptoms, accurately identifying HLB samples at each symptom stage remains a challenge.

[0004] Currently, the main methods for detecting Huanglongbing (HLB) include field diagnosis, microscopic observation, molecular biology methods, volatile organic compound (VOC) detection, and imaging spectroscopy. Field diagnosis relies on experienced technicians, is highly subjective, and has a high false-positive rate, making it only suitable for preliminary screening. Microscopic observation depends on sample preparation and slide selection, and can be used as an auxiliary research method, but is not suitable for large-scale field testing. Molecular biology methods, mainly PCR, are highly sensitive, but the equipment is expensive and the operation is cumbersome, making it difficult to promote on a large scale in production areas. VOC detection utilizes the analysis of volatile substances produced by plant stress; studies have shown that specific volatile substances can be produced in the early stages of HLB infection, aiding in early diagnosis, but the commonly used large instruments such as GC-MS are also complex to operate and costly, making widespread adoption difficult. Imaging spectroscopy analyzes the spectra and images of citrus leaves or plants, combined with machine learning for detection, but early HLB symptoms are not obvious, and late-stage zinc deficiency symptoms overlap with HLB symptoms, easily leading to missed detections and misdiagnosis.

[0005] In existing technologies, detection methods based on single imaging (such as multi-model fusion and attention-enhanced YOLO networks) have improved accuracy in complex environments, but they struggle to capture internal physicochemical features and are easily affected by similar symptoms such as zinc deficiency. Visible-near-infrared spectroscopy-based methods can detect internal chemical changes but lack leaf texture information. Early attempts at multi-modal fusion (such as feature splicing) have shown potential for integrating multi-source data, but their fusion strategies are simple and lack the ability to dynamically weight cross-modal features. Furthermore, electronic nose-based detection methods show good performance in identifying typical Huanglongbing symptoms, but their ability to identify early-stage asymptomatic or late-stage nutrient-deficient samples is limited. Most existing studies employ single-sensor technologies, which cannot comprehensively acquire the host's response characteristics to Huanglongbing stress, making it difficult to simultaneously meet the accurate identification requirements of early, mid-stage typical symptom, and late-stage zinc-deficient Huanglongbing samples.

[0006] In summary, current methods for detecting Huanglongbing (HLB) are limited by single-sensor technology, making it difficult to accurately identify samples at various symptom stages. There is an urgent need for a rapid detection method that can integrate multi-source sensor information, cover all infection stages of HLB, and achieve high detection accuracy. Solving these technical problems is the challenge addressed by this invention. Summary of the Invention

[0007] The purpose of this invention is to provide a method for detecting Huanglongbing (HLB) in citrus based on the fusion of visual and olfactory information. This method addresses the technical problems in existing technologies, such as the inability of single-sensor technology to accurately identify different stages of HLB symptoms, missed detections and misjudgments of early HLB and zinc-deficient HLB samples, and insufficient fusion of multimodal features. By fusing visual and olfactory features at the feature layer, this invention achieves accurate identification of six types of samples: healthy, magnesium-deficient, zinc-deficient, zinc-deficient HLB, early HLB, and HLB with typical symptoms.

[0008] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0009] A method for detecting Huanglongbing (HLB) in citrus fruits based on the fusion of visual and olfactory information employs a computer vision device and an electronic nose device. The computer vision device integrates two modes: reflectance imaging and transmission imaging. The reflectance image uses a 660nm single-band light source and multi-angle polarization imaging to acquire leaf surface texture information, while the transmission image uses a 590nm band light source transmitted from the bottom of the leaf and collected by a polarization camera to obtain information on starch and soluble sugar accumulation inside the leaf. The electronic nose device includes a gas collection chamber, a test chamber, and a data acquisition module. The test chamber consists of an inlet chamber, a reaction chamber, and an outlet chamber, and uses a sensor array of eight gas-sensitive sensors to collect the characteristics of volatile organic compounds in the leaves through dynamic intake and exhaust.

[0010] The detection method includes the following steps:

[0011] Step 1: Collect citrus leaf samples. Take the main vein of the leaf for qPCR to confirm its health status, and use the remaining leaf tissue for electronic nose detection. First, use a computer vision device to collect leaf reflection and transmission images, and then use an electronic nose device to collect volatile gases from the leaves. The collection temperature of the electronic nose device is 40℃, the collection time is 10min, and the sample weight is 0.2g.

[0012] Step 2: Preprocess the reflected image and extract visual features. For the transmitted image, preprocess the reflected image and extract visual features. When preprocessing the reflected image, select the 90° polarized image as the analysis object, and sequentially perform grayscale conversion, threshold segmentation, region extraction, background removal, ellipse fitting, and affine transformation. When preprocessing the transmitted image, calculate the linear polarization angle (AoLP) image and perform image enhancement. The visual features include grayscale co-occurrence matrix texture features and grayscale histogram features. The texture features include the mean and standard deviation of four parameters—energy, contrast, correlation, and inverse difference—in the four directions of 0°, 45°, 90°, and 135°. The grayscale histogram features include the grayscale mean and grayscale standard deviation.

[0013] Step 3: Extract features from the electronic nose sensor response signal to obtain olfactory features; the olfactory features are the extreme value features of the sensor response curve, that is, the sensor response value at the peak or trough of the signal curve;

[0014] Step 4: The visual and olfactory features are fused using a feature layer fusion method. The feature layer fusion method adopts an improved recursive feature elimination and cross-validation (RFECV) method. The RFECV algorithm is used to select features for each classification step in the stepwise classification model. The optimal feature subset and the optimal classification model for each step are determined by five-fold cross-validation. The fused features include electronic nose volatile organic compound features, reflectance imaging features, and transmission imaging features.

[0015] Step 5: Establish a stepwise classification model, select the optimal classifier for each classification step, and distinguish between six types of samples: healthy, magnesium-deficient, zinc-deficient, zinc-deficient Huanglongbing (HLB), early-stage HLB, and typical-symptom HLB. The stepwise classification model includes five sequential classification stages: First, divide the samples into two major categories based on the presence or absence of visible yellowing symptoms; Second, distinguish between healthy and early-stage HLB based on the presence or absence of starch and soluble sugar accumulation; Third, screen for the zinc-deficient category from the four leaf types based on leaf texture differences; Fourth, distinguish between zinc-deficient and zinc-deficient HLB based on the degree of starch and soluble sugar accumulation; Fifth, distinguish between magnesium-deficient and typical-symptom HLB based on differences in yellowing texture. Steps 1, 3, and 5 use reflectance imaging features, while steps 2 and 4 use electronic nose olfactory features. The optimal classifier is determined by comparing the performance of Linear Discriminant Analysis (LDA), Random Forest (RF), Support Vector Machine (SVM), Classification Regression Tree (CART), and Logistic Regression (LR) in terms of accuracy, precision, recall, and F1 score.

[0016] Preferably, the computer vision device is a self-made portable device that integrates reflection imaging and transmission imaging, and uses a polarization camera to acquire polarization images at four relative angles of 0°, 45°, 90°, and 135°.

[0017] Preferably, the air inlet chamber of the electronic nose device is cone-shaped, and the air outlet chamber is hemispherical, employing a dynamic air intake method with one inlet and one outlet. The sensor is a silicon dioxide type gas sensor, and its model and corresponding target gas are TGS822, TGS822TF, TGS826, TGS2600, MQ3B, MQ136, MQ138, and WSP1110.

[0018] Preferably, in step four, the optimal classification model for each classification step is selected from linear discriminant analysis (LDA) or logistic regression (LR).

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0020] (1) This invention adopts visual and olfactory dual-modal sensor information fusion. It integrates the leaf surface texture and internal starch accumulation information obtained by computer vision device with the volatile organic compound characteristics obtained by electronic nose device at the feature layer, which makes up for the shortcomings of single sensor source in the identification of early Huanglongbing, zinc deficiency Huanglongbing and healthy and nutrient deficiency samples, and realizes the complementary fusion of multi-source sensor information.

[0021] (2) The present invention adopts an improved RFECV feature selection method, which performs recursive feature elimination and cross-validation on visual and olfactory features in each classification step of the stepwise classification model, and selects the optimal feature subset and optimal classifier in each step, thereby overcoming the defects of traditional direct fusion or simple stepwise fusion and improving the stability and accuracy of the model in identifying Huanglongbing samples in complex environments.

[0022] (3) This invention achieves accurate identification of samples at various symptom stages of Huanglongbing. The overall accuracy rate of identifying positive Huanglongbing samples is 95.38%, the accuracy rate of identifying early Huanglongbing samples is 94.23%, and the accuracy rate of identifying zinc-deficient Huanglongbing samples is 94.12%. It is suitable for rapid on-site detection of Huanglongbing in citrus and some nutrient deficiency symptoms, and provides referable technical support for the application of multi-source sensor information and multi-modal feature fusion of plant diseases. Attached Figure Description

[0023] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0024] Figure 1 This is a flowchart of a method for detecting Huanglongbing (HLB) in citrus based on the fusion of visual and olfactory information.

[0025] Figure 2 Color images of different types of citrus leaf samples.

[0026] Figure 3 (a) is a schematic diagram of the self-made computer vision device, (b) is a physical image of the self-made computer vision device, and (c) is an example of the collected reflected and transmitted images.

[0027] Figure 4 This describes the process of extracting visual features from images acquired using a self-made computer vision system.

[0028] Figure 5 This is a schematic diagram of a stepwise classification model based on a self-made computer vision device.

[0029] Figure 6 (a) shows a comparison of airflow simulations for different structures of the inlet and outlet chambers; (b) shows a comparison of airflow simulations for different positions of the exhaust port in the reaction chamber; (c) shows a schematic diagram of the electronic nose test chamber; and (d) shows a schematic diagram of the process of testing samples using a self-made electronic nose device.

[0030] Figure 7 The images show the actual components of the homemade electronic nose system.

[0031] Figure 8 This is a heatmap showing the average accuracy of the RFECV algorithm based on five-fold cross-validation of the training set. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0033] Example 1: This example provides a method for detecting Huanglongbing in citrus based on the fusion of visual and olfactory information, using a self-made computer vision device and a self-made electronic nose device.

[0034] like Figure 3 As shown, the computer vision device is a self-made portable device integrating reflection and transmission imaging. It uses a polarization camera to acquire polarized images at four relative angles: 0°, 45°, 90°, and 135°. The reflection images use a 660nm single-band light source, highlighting the yellowing areas of the leaves based on the chlorophyll absorption spectrum characteristics, and employ multi-angle polarization imaging to eliminate interference from the reflection of the leaf's waxy layer. The transmission images use a 590nm band light source transmitted from the bottom of the leaf, and the transmitted light is acquired by the polarization camera to obtain information on starch and soluble sugar accumulation inside the leaf. Color images of different types of citrus leaf samples are shown below. Figure 2 As shown, the samples include six categories: healthy, magnesium-deficient, zinc-deficient, zinc-deficient Huanglongbing, early-stage Huanglongbing, and Huanglongbing with typical symptoms.

[0035] like Figure 6 and Figure 7 As shown, the electronic nose device includes an air collection chamber, a testing module, and a data acquisition module. The testing chamber consists of an air inlet chamber, a reaction chamber, and an air outlet chamber, employing a dynamic air intake and exhaust method with one inlet and one outlet. The air inlet chamber is cone-shaped, and the air outlet chamber is hemispherical. The sensor array consists of eight silicon dioxide gas sensors, model numbers TGS822, TGS822TF, TGS826, TGS2600, MQ3B, MQ136, MQ138, and WSP1110.

[0036] like Figure 1 As shown, the detection method includes the following steps:

[0037] Step 1: Sample Collection and Data Acquisition

[0038] Leaf samples were collected from the Ponkan orange orchard of the Zhejiang Provincial Citrus Research Institute. Citrus plants were categorized into those infected with Huanglongbing (HLB) or suffering from nutrient deficiency, based on expert advice. Leaves were collected from three trees in each category, collected from each tree from all four directions (north, south, east, and west). After collection, the leaves were immediately transported to the laboratory via cold chain and stored at 4℃. First, reflectance and transmission images of the leaves were acquired using a computer vision device. Then, the midrib of the leaves was sampled for real-time quantitative PCR (qPCR) to confirm their health status. Finally, the remaining leaf tissue was used for electronic nose detection. The electronic nose gas collection temperature was 40℃, the collection time was 10 min, and the sample weight was 0.2 g. 0.2 g of the sample was placed in a 200 mL gas collection bottle, sealed, and placed in a 40℃ constant temperature gas collection chamber for 10 min.

[0039] Step 2: Visual Feature Extraction

[0040] like Figure 4 As shown, when preprocessing the reflected image, a 90° polarized image is selected as the analysis object based on the anti-reflection effect. The following steps are performed sequentially: (a) grayscale conversion, (b) threshold segmentation, (c) region extraction, (d) background removal, (e) ellipse fitting, and (f) affine transformation, to obtain a blade image with consistent orientation and no background. When preprocessing the transmitted image, the process includes: (i) multi-angle polarization image acquisition, (ii) calculation of the linear polarization angle (AoLP) image, and (iii) image enhancement.

[0041] The polarization angle characterizes the angular relationship between the direction of light vibration and a reference direction. When polarized light passes through an optically active substance, the plane of vibration rotates, causing a change in the polarization angle. The formula for calculating AoLP is:

[0042] ;

[0043] ;

[0044] in, Total light intensity The difference in intensity between linearly polarized light at 0° and 90°. The difference in intensity between linearly polarized light at 45° and 135° directions. This represents the intensity difference between right-handed and left-handed circularly polarized light.

[0045] Because the AoLP images of leaves with low starch content have weak contrast, a robust image enhancement method is used. The grayscale transformation formula is:

[0046] ;

[0047] ;

[0048] ;

[0049] ;

[0050] ;

[0051] in, The initial grayscale value. , To enhance the coefficient, These are the scaled grayscale values. , The maximum and minimum grayscale values ​​of the image. The grayscale value is The number of pixels, This represents the total number of pixels within the region of interest in the image. grayscale value The cumulative probability density.

[0052] The preprocessed image undergoes texture and gray-level histogram feature extraction. Texture features are based on the Gray-Level Co-occurrence Matrix (GLCM), with gray levels quantized to 256 levels and a sampling distance d=1. The GLCM is calculated in four directions: 0°, 45°, 90°, and 135°, and four texture features—energy, correlation, inverse difference, and contrast—are extracted.

[0053] ;

[0054] ;

[0055] ;

[0056] ;

[0057] in, For gray levels, , Grayscale value for Probability of occurrence , These represent the mean and variance of the matrix. Gray-level histogram features include the gray-level mean and standard deviation:

[0058] ; ;

[0059] in, For region of interest, For pixels, For pixel grayscale values, This represents the total number of pixels. Ten-dimensional visual features are obtained from both the reflected and transmitted images.

[0060] Step 3: Extraction of olfactory features

[0061] Extreme value features, i.e., the sensor response values ​​at the peaks or troughs of the signal curve, were extracted from the response curve of the electronic nose sensor to characterize the composition and concentration of volatile organic compounds and the sensor's response intensity to different volatile substances. The inlet flow rate was set to 300 mL / min, the outlet flow rate to 30 mL / min, the sensor cleaning time to 180 s, the measurement time to 300 s, and the sampling interval to 1 s. Each sensor collected 300 data points, and the total of 2400 data points from the 8 sensors represented one sample.

[0062] Step 4: Fusion of visual and olfactory features

[0063] Feature layer fusion was employed, utilizing a total of 28 feature variables from both visual and olfactory perspectives: 10 features from reflectance imaging, 10 features from transmission imaging, and 8 features from volatile organic compound (VOC) sensors. The feature sources and dimensions of the direct fusion method are shown in Table 1.

[0064] Table 1. Feature sources and variable dimensions used in the direct fusion method

[0065]

[0066] Feature selection is performed using an improved RFECV method: In such cases... Figure 5 The stepwise classification model shown uses the RFECV algorithm in each classification step, and determines the optimal feature subset and optimal classification model for each step through five-fold cross-validation. The heatmap of the average accuracy of RFECV five-fold cross-validation is shown below. Figure 8 As shown in Table 2, the RFECV feature selection results are as follows.

[0067] Table 2 Feature selection results based on RFECV

[0068]

[0069] The feature numbers are arranged in the order of olfactory features, reflectance image features, and transmission image features in Table 1. Features 0–7 are obtained from the 8 sensors of the electronic nose, features 8–17 are obtained from the 10 features acquired from the reflectance image, and features 18–27 are obtained from the 10 features acquired from the transmission image. The fused feature vector set obtained by improving RFECV feature selection is shown in Table 3.

[0070] Table 3. Fusion Feature Vector Dataset Based on Improved RFECV Feature Selection Method

[0071]

[0072] Step 5: Step-by-step classification and identification

[0073] A five-step stepwise classification model was established: Step 1, samples were divided into two main categories based on the presence or absence of visible yellowing symptoms; Step 2, healthy samples were distinguished from early-stage Huanglongbing (HLB) based on the presence or absence of starch and soluble sugar accumulation; Step 3, zinc deficiency was screened from the four leaf categories based on differences in leaf texture; Step 4, zinc deficiency was distinguished from zinc-deficient HLB based on the degree of starch and soluble sugar accumulation; Step 5, magnesium deficiency was distinguished from HLB with typical symptoms based on differences in yellowing texture. Steps 1, 3, and 5 used reflectance imaging features, while steps 2 and 4 used electronic nose olfactory features. The optimal classifier was selected by comparing the accuracy, precision, recall, and F1 score of LDA, RF, SVM, CART, and LR. The following evaluation metrics were calculated using a confusion matrix:

[0074] ;

[0075] ;

[0076] ;

[0077] ;

[0078] Wherein, TP represents true positive, which is the number of samples that are actually positive for Huanglongbing and are correctly identified as positive for Huanglongbing; TN represents true negative, which is the number of samples that are actually negative for Huanglongbing and are correctly identified as negative for Huanglongbing; FP represents false positive, which is the number of samples that are actually negative for Huanglongbing but are misclassified as positive for Huanglongbing; FN represents false negative, which is the number of samples that are actually positive for Huanglongbing but are missed being classified as negative for Huanglongbing; P represents precision, which represents the proportion of samples that the model classifies as positive for Huanglongbing that are actually positive for Huanglongbing; and R represents recall, which represents the proportion of samples that are actually positive for Huanglongbing that are correctly identified.

[0079] Example 2:

[0080] Based on Example 1, a classification model established using the improved RFECV feature selection method was used to detect 390 leaf samples. The identification results of the six types of samples are shown in Table 4. The overall identification accuracy rate was 92.31%, and the identification rate of each of the six types of samples was around 90%.

[0081] Table 4. Confusion Matrix of Test Results for Feature Set Screening Based on Improved RFECV Method

[0082]

[0083] The samples were divided into Huanglongbing (HLB) positive (early HLB, typical symptom HLB, zinc deficiency HLB) and HLB negative (healthy, magnesium deficiency, zinc deficiency) according to their HLB infection status. The confusion matrix is ​​shown in Table 5. The accuracy rate of HLB positive sample identification was 95.38%.

[0084] Table 5. Classification results of Huanglongbing infected samples based on the improved RFECV feature selection method.

[0085]

[0086] When using computer vision features combined with a stepwise classification model to identify six types of samples, the overall recognition rate was 82.31%, while the recognition rates for zinc deficiency and early-stage Huanglongbing (HLB) were below 80%. After employing an improved RFECV method that integrates visual and olfactory features, the accuracy rate for identifying HLB-positive samples increased from 84.62% with computer vision alone and 93.59% with an electronic nose alone to 95.38%. The recognition rate for early-stage HLB reached 94.23%, and the recognition rate for zinc deficiency HLB reached 94.12%, verifying the effectiveness and practicality of the method of this invention.

[0087] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting citrus Huanglongbing (HLB) based on the fusion of visual and olfactory information, characterized in that, The device employs a computer vision device and an electronic nose device; the computer vision device integrates both reflective and transmissive imaging modes; the electronic nose device includes an air collection chamber, a testing chamber, and a data acquisition module. The detection method includes the following steps: Step 1: Collect citrus leaf samples. Take the main vein of the leaf for qPCR to confirm the health status. Use the remaining leaf tissue for electronic nose detection. First, use a computer vision device to collect leaf reflection and transmission images, and then use an electronic nose device to collect leaf volatile gases. Step 2: Preprocess the reflected image and extract visual features; preprocess the transmitted image and extract visual features. Step 3: Extract features from the electronic nose sensor response signal to obtain olfactory characteristics; Step 4: Use a feature layer fusion method to fuse visual features and olfactory features; Step 5: Establish a stepwise classification model, select the optimal classifier for each classification step, and use it to distinguish between six types of samples: healthy, magnesium-deficient, zinc-deficient, zinc-deficient Huanglongbing, early Huanglongbing, and Huanglongbing with typical symptoms.

2. The method for detecting citrus Huanglongbing based on the fusion of visual and olfactory information according to claim 1, characterized in that, In step one, the reflected image uses a 660nm single-band light source and multi-angle polarization imaging; the transmitted image uses a 590nm band light source transmitted from the bottom of the blade, and the transmitted light is collected by a polarization camera to form an image; the air collection temperature of the electronic nose device is 40℃, the air collection time is 10min, and the sample weight is 0.2g.

3. The method for detecting citrus Huanglongbing based on the fusion of visual and olfactory information according to claim 1, characterized in that, In step two, when preprocessing the reflected image, a 90° polarized image is selected as the analysis object, and grayscale conversion, threshold segmentation, region extraction, background removal, ellipse fitting, and affine transformation are performed sequentially; when preprocessing the transmitted image, the linear polarization angle AoLP image is calculated and image enhancement is performed. The formula for calculating AoLP is: ； ； in, Total light intensity The difference in intensity between linearly polarized light at 0° and 90°. The difference in intensity between linearly polarized light at 45° and 135° directions. This represents the intensity difference between right-handed and left-handed circularly polarized light. Using a robust image enhancement method, the grayscale value transformation formula is as follows: ； ； ； ； ； in, The initial grayscale value. , To enhance the coefficient, These are the scaled grayscale values. , The maximum and minimum grayscale values ​​of the image. The grayscale value is The number of pixels, This represents the total number of pixels within the region of interest in the image. grayscale value The cumulative probability density; The visual features include gray-level co-occurrence matrix texture features and gray-level histogram features. The texture features include the mean and standard deviation of four parameters: energy, contrast, correlation, and inverse difference in four directions: 0°, 45°, 90°, and 135°.

4. The method for detecting citrus Huanglongbing based on the fusion of visual and olfactory information according to claim 3, characterized in that, The formulas for calculating the four texture features—energy, correlation, inverse difference, and contrast—are as follows: ； ； ； ； in, For gray levels, , Grayscale value for Probability of occurrence , These are the mean and variance of the matrix; Gray-level histogram features include the gray-level mean and gray-level standard deviation: ； ； in, For region of interest, For pixels, For pixel grayscale values, The total number of pixels; 10-dimensional visual features are obtained from both the reflected and transmitted images.

5. The method for detecting citrus Huanglongbing based on the fusion of visual and olfactory information according to claim 1, characterized in that, In step three, the olfactory feature is the extreme value feature of the electronic nose sensor response curve, which is the sensor response value at the peak or trough of the signal curve.

6. The method for detecting citrus Huanglongbing based on the fusion of visual and olfactory information according to claim 1, characterized in that, In step four, the feature layer fusion method adopts an improved recursive feature elimination and cross-validation (RFECV) method; the RFECV algorithm is used to select features for each classification step in the stepwise classification model, and the optimal feature subset and optimal classification model for each step are determined through cross-validation.

7. The method for detecting citrus Huanglongbing based on the fusion of visual and olfactory information according to claim 6, characterized in that, The RFECV method employs five-fold cross-validation; the fusion features include electronic nose volatile organic compound features, reflectance imaging features, and transmission imaging features; The optimal classification model for each classification step is selected from linear discriminant analysis (LDA) or logistic regression (LR).

8. The method for detecting citrus Huanglongbing based on the fusion of visual and olfactory information according to claim 1, characterized in that, In step five, the stepwise classification model includes the following five sequential classification stages: The first step is to divide the samples into two main categories based on the presence or absence of visible yellowing symptoms; The second step is to differentiate between healthy individuals and those in the early stages of Huanglongbing based on the presence or absence of starch and soluble sugar accumulation. The third step is to screen for zinc deficiency categories from the four types of leaves based on differences in leaf texture. The fourth step is to differentiate between zinc deficiency and zinc-deficiency-type Huanglongbing based on the degree of starch and soluble sugar accumulation. The fifth step is to differentiate between magnesium deficiency and Huanglongbing (HLB), a disease characterized by typical symptoms, based on the differences in yellowing patterns. The first, third, and fifth steps utilize reflection imaging features, while the second and fourth steps utilize electronic nose olfactory features.

9. The method for detecting citrus Huanglongbing based on the fusion of visual and olfactory information according to claim 1, characterized in that, In step five, the optimal classifier is determined by comparing the performance of Linear Discriminant Analysis (LDA), Random Forest (RF), Support Vector Machine (SVM), Classification and Regression Tree (CART), and Logistic Regression (LR) in terms of accuracy, precision, recall, and F1 score. The following evaluation metrics were calculated using a confusion matrix: ； ； ； ； Wherein, TP represents true positive, which is the number of samples that are actually positive for Huanglongbing and are correctly identified as positive for Huanglongbing; TN represents true negative, which is the number of samples that are actually negative for Huanglongbing and are correctly identified as negative for Huanglongbing; FP represents false positive, which is the number of samples that are actually negative for Huanglongbing but are misclassified as positive for Huanglongbing; FN represents false negative, which is the number of samples that are actually positive for Huanglongbing but are missed being classified as negative for Huanglongbing; P represents precision, which represents the proportion of samples that the model classifies as positive for Huanglongbing that are actually positive for Huanglongbing; and R represents recall, which represents the proportion of samples that are actually positive for Huanglongbing that are correctly identified.

10. The method for detecting citrus Huanglongbing based on the fusion of visual and olfactory information according to claim 1, characterized in that, The computer vision device uses a polarization camera to acquire polarization images at four relative angles: 0°, 45°, 90°, and 135°. The electronic nose device includes a sensor array of eight gas-sensitive sensors. The test chamber consists of an air inlet chamber, a reaction chamber, and an air outlet chamber, and adopts a dynamic air intake and exhaust method.

Images