Tobacco leaf oil content characterization method based on image processing and volatile metabolite analysis

Through image processing and volatile metabolites analysis, combined with U-Net and GC-MS technology, the CARS-RF model was established, which solved the problem of low efficiency and poor accuracy of tobacco leaf oil content evaluation in the existing technology, and achieved accurate evaluation of tobacco leaf oil content quality and efficient prediction.

CN119068245APending Publication Date: 2024-12-03HENAN AGRICULTURAL UNIVERSITY
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411134603.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-19
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

In the prior art, the evaluation of tobacco leaves oil content mainly relies on manual observation, which is inefficient and poorly accurate, making it difficult to accurately evaluate the appearance quality of tobacco leaves.

Method used

Using a method based on image processing and volatile metabolites analysis, the background segmentation of tobacco leaves is performed through the U-Net convolution network, combined with GC-MS volatile metabolomics analysis, and the feature selection method is used to extract images and metabolites characteristics to establish CARS -RF fusion model performs oil classification prediction.

Benefits of technology

The accurate evaluation of the quality of tobacco oil components was achieved, the evaluation efficiency and accuracy were improved, the relative importance of outputting various influencing factors was established, and a tobacco oil component quality prediction model based on images and metabolites was established.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068245A_ABST
    Figure CN119068245A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of tobacco leaf oil content characterization, in particular to a tobacco leaf oil content characterization method based on image processing and volatile metabolite analysis, which comprises the following steps: placing tobacco leaves in a still life box to shoot images, and analyzing metabolite composition of tobacco leaves with different oil content categories by adopting GC-MS volatile metabonomics; analyzing the metabolite abundance difference of the tobacco leaves with different oil grades; carrying out background segmentation on the collected tobacco leaf image by adopting a U-Net convolutional network; tobacco leaf image feature information is obtained, the tobacco leaf image information is extracted through an Ai seed scanning system, and image features are extracted. The method comprises the following steps: representing tobacco leaf oil content quality by using appearance image features and volatile metabolite abundance, extracting original high-latitude data information, carrying out feature variable screening by using a CARS algorithm, establishing an oil content grade prediction model in combination with an RF algorithm, and organically fusing feature selection with the prediction model to obtain a prediction result. And establishing a tobacco leaf oil content quality prediction model based on a CARS-RF fusion model, and outputting relative importance of various influence factors.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of tobacco leaf oil characterization, and in particular to a tobacco leaf oil characterization method based on image processing and volatile metabolite analysis. Background Art

[0002] The appearance quality of tobacco leaves includes factors such as maturity, color, oil content, identity, structure, and chromaticity. Oil content is an important basis for reflecting the appearance quality of tobacco leaves and plays an important role in evaluating the industrial usability of tobacco leaves. Oil is a soft semi-liquid or liquid substance that exists in the tissue cells of tobacco leaves. Its appearance gives people a feeling of oiliness and fullness, and is closely related to the appearance quality, chemical composition, and sensory quality of tobacco leaves.

[0003] The existing tobacco leaf quality is generally assessed by manual observation, which makes it difficult to accurately assess the oil content of tobacco leaves. In addition, human subjective judgment has disadvantages such as low efficiency and poor accuracy. Summary of the invention

[0004] In view of the above-mentioned shortcomings of the prior art, the present invention provides a tobacco leaf oil characterization method based on image processing and volatile metabolite analysis, which can effectively solve the problems of low efficiency and poor accuracy of human subjective judgment in the prior art.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0006] The present invention provides a tobacco leaf oil characterization method based on image processing and volatile metabolite analysis, comprising the following steps:

[0007] S100: Place tobacco leaves in a still box and take images. Use image processing technology to extract information such as morphology, color, and texture in the images. Use GC-MS volatile metabolomics to analyze the metabolite composition of tobacco leaves of different oil categories, analyze the differences in metabolite abundance of tobacco leaves of different oil grades, and use feature selection methods to mine key images and metabolite features that can characterize the oil quality of tobacco leaves.

[0008] S200: The U-Net convolutional network is used to perform background segmentation on the collected tobacco leaf image. The U-Net network structure consists of two parts: the first part is the feature extraction part on the left, and the second part is the upsampling part on the right;

[0009] S300: Acquire tobacco leaf image feature information, use the Ai seed scanning system to extract tobacco leaf image information, and extract image features, wherein the image features include morphological feature values, color feature values, and texture feature values.

[0010] Further, S110: weighing a tobacco leaf sample, grinding it and placing it in a headspace bottle to obtain a powder sample;

[0011] S120: adding a saturated NaCl solution and an internal standard solution to the powder sample to obtain a mixed solution at a constant temperature of 60° C.;

[0012] S130: The mixed solution is subjected to headspace extraction using a fully automatic headspace solid phase microextraction device for GC-MS separation and identification;

[0013] S140: Establish a database, select 1 quantitative ion and 2-3 qualitative ions for each compound and perform precise scanning. Each group of detected ions is detected in the order of peaks and in different time periods. If the detected retention time is consistent with the standard reference, and the selected ions appear in the sample mass spectrum after background subtraction, it is determined to be the substance, and the quantitative ion is selected for integration and correction to enhance quantitative accuracy. At the same time, based on the original data, the missing values ​​are first filled using the k nearest neighbor algorithm, and then the QC sample CV value is calculated, and the substances with CV values ​​less than 0.5 are retained to obtain the final data. In the extraction test, 10 biological replicates were performed for each grade of oil sample detection.

[0014] Furthermore, the MassHunter software was used to process the mass spectrometry files for integration and correction. An isotope internal standard was selected and the relative content of VOCs in the sample was calculated according to the following formula;

[0015] The formula for calculating the relative content of compounds in solid samples is:

[0016]

[0017] X i is the content of compound i in the sample to be tested (μg / g); V s is the volume of internal standard added (μL); C s is the concentration of the internal standard (μg / mL); M is the amount of the sample to be tested (g); I s is the peak area of ​​the internal standard; I i is the peak area of ​​compound i in the sample to be tested;

[0018] The formula for calculating the relative content of compounds in liquid samples is:

[0019]

[0020] X i is the content of compound i in the sample to be tested (μg / mL); V s is the volume of internal standard added (μL); C s is the concentration of the internal standard (μg / mL); V is the volume of the sample to be tested (mL); I s is the peak area of ​​the internal standard; I i is the peak area of ​​compound i in the sample to be tested.

[0021] Furthermore, orthogonal partial least squares discriminant analysis was used to decompose X into two types of information related to and unrelated to Y, and the difference variables were screened by removing the differences of the unrelated orthogonal principal component information;

[0022] After OPLS-DA performs log2 transformation on the original data, it performs central processing and calculates the VIP value of the metabolite variable importance projection. The VIP value represents the influence of the inter-group difference of the corresponding metabolite on the classification and discrimination of each group of samples in the model.

[0023] VIP>1 means significant difference, and the difference fold value is combined to further screen out the differential metabolites between different samples; metabolites with VIP>1 are screened and selected, and then metabolites with fo ld change ≥ 2 and fo ld change ≤ 0.5 are selected. Metabolites with a difference of more than 2 times or less than 0.5 between the control group and the treatment group are considered differential metabolites.

[0024] Furthermore, the random forest algorithm was used to construct the oil grade prediction model, and the selected image and metabolite features were modeled separately.

[0025] Compared with the known prior art, the technical solution provided by the present invention has the following beneficial effects:

[0026] The appearance image features and the abundance of volatile metabolites are used to characterize the oil quality of tobacco leaves. The original high-dimensional data information is extracted, and the CARS algorithm is used to screen feature variables. The RF algorithm is combined to establish an oil grade prediction model. The feature selection and prediction model are organically integrated to establish a tobacco leaf oil quality prediction model based on the CARS-RF fusion model, and the relative importance of various influencing factors is output. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0028] Figure 1 This is a schematic diagram of tobacco leaf image segmentation based on U-Net network;

[0029] Figure 2 for Figure 1 Information diagram of tobacco leaf image features;

[0030] Figure 3 Build a process for predicting tobacco leaf oil content;

[0031] Figure 4 The image feature differences of tobacco leaves with different oil content;

[0032] Figure 5 To analyze the differential metabolites among different oil groups;

[0033] Figure 6 For KEGG pathway enrichment analysis;

[0034] Figure 7 It is a radar chart of sensory flavor characteristics;

[0035] Figure 8 It is a diagram of the CARS feature screening process;

[0036] Fig. 9 It is the key feature correlation analysis diagram;

[0037] Fig.10 This is a feature importance analysis diagram;

[0038] Fig.11 This is a prediction chart for tobacco leaf oil quality. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] The present invention will be further described below in conjunction with the embodiments.

[0041] Embodiment: A method for characterizing tobacco leaf oil based on image processing and volatile metabolite analysis comprises the following steps:

[0042] S100: Place tobacco leaves in a still box and take images. Use image processing technology to extract information such as morphology, color, and texture in the images. Use GC-MS volatile metabolomics to analyze the metabolite composition of tobacco leaves of different oil categories, analyze the differences in metabolite abundance of tobacco leaves of different oil grades, and use feature selection methods to mine key images and metabolite features that can characterize the oil quality of tobacco leaves.

[0043] S200: The U-Net convolutional network is used to perform background segmentation on the collected tobacco leaf image. The U-Net network structure consists of two parts: the first part is the feature extraction part on the left, and the second part is the upsampling part on the right;

[0044] like Figure 2As shown in the figure, the U-Net convolutional network is used to segment the background of the original image. The U-Net network structure consists of two parts, the first part is the feature extraction part on the left, and the second part is the upsampling part on the right. Because its network structure is similar to the letter U, it is called the U-Net network. In the feature extraction part on the left, each time a pooling layer is passed, a feature map of a new scale will be generated, and there are 5 scales in total with the original image. In the upsampling part on the right, each time an upsampling is passed, a feature map corresponding to the scale of the left part will be generated, and it will be spliced ​​with the feature map on the left. After 5 layers of downsampling and 5 layers of upsampling, two layers are finally output, which are the foreground and background. In this experiment, a total of 601 tobacco leaf image samples were collected, and the U-Net convolutional network was used for training. They were randomly divided into training sets and test sets in a ratio of 7:3 to achieve semantic segmentation of tobacco leaf images.

[0045] S300: Acquire tobacco leaf image feature information, use the Ai seed scanning system to extract tobacco leaf image information, and extract image features, wherein the image features include morphological feature values, color feature values, and texture feature values.

[0046] The tobacco leaf image information was extracted using the Ai seed (Nanjing Zhinong Yunxin Big Data Technology Co., Ltd.) scanning system, and 54 image features (morphology, color, and texture) were extracted, including 6 morphological feature values, 20 color feature values, and 28 texture feature values.

[0047] S110: weighing a tobacco leaf sample, grinding it and placing it in a headspace bottle to obtain a powder sample;

[0048] S120: adding a saturated NaCl solution and an internal standard solution to the powder sample to obtain a mixed solution at a constant temperature of 60° C.;

[0049] S130: The mixed solution is subjected to headspace extraction using a fully automatic headspace solid phase microextraction device for GC-MS separation and identification;

[0050] S140: Establish a database, select 1 quantitative ion and 2-3 qualitative ions for each compound and perform precise scanning. Each group of detected ions is detected in the order of peaks and in different time periods. If the detected retention time is consistent with the standard reference, and the selected ions appear in the sample mass spectrum after background subtraction, it is determined to be the substance, and the quantitative ion is selected for integration and correction to enhance quantitative accuracy. At the same time, based on the original data, the missing values ​​are first filled using the k nearest neighbor algorithm, and then the QC sample CV value is calculated, and the substances with CV values ​​less than 0.5 are retained to obtain the final data. In the extraction test, 10 biological replicates were performed for each grade of oil sample detection.

[0051] 0.5 g tobacco leaf sample was weighed and ground, placed in a headspace bottle, and saturated NaCl solution and 20 μL (10 μg / mL) internal standard solution were added. The headspace was extracted for 15 min using a fully automatic headspace solid phase microextraction device (SPME Arrow, CTC Analyt i cs AG) at a constant temperature of 60 ° C for GC-MS separation and identification. Based on the self-built database, the selected ion detection mode was used for precise scanning. One quantitative ion and 2 to 3 qualitative ions were selected for each compound. Each group of detected ions was detected in different time periods according to the peak order. If the detected retention time was consistent with the standard reference, and the selected ions appeared in the sample mass spectrum after background subtraction, it was determined to be the substance, and the quantitative ion was selected for integration and correction to enhance the quantitative accuracy. At the same time, based on the original data, the k-nearest neighbors algorithm (KNN) was used to fill the missing values, and then the QC sample CV value was calculated, and the substances with CV values ​​less than 0.5 were retained to obtain the final data. In the extraction experiment, 10 biological replicates were performed for each grade of oil sample.

[0052] Use MassHunter software to process the mass spectrometry files, perform integration and correction. Select an isotope internal standard and calculate the relative content of VOCs in the sample according to the following formula;

[0053] The formula for calculating the relative content of compounds in solid samples is:

[0054]

[0055] X i is the content of compound i in the sample to be tested (μg / g); V s is the volume of internal standard added (μL); C s is the concentration of the internal standard (μg / mL); M is the amount of the sample to be tested (g); I s is the peak area of ​​the internal standard; I i is the peak area of ​​compound i in the sample to be tested;

[0056] The formula for calculating the relative content of compounds in liquid samples is:

[0057]

[0058] X i is the content of compound i in the sample to be tested (μg / mL); V s is the volume of internal standard added (μL); C s is the concentration of the internal standard (μg / mL); V is the volume of the sample to be tested (mL); I s is the peak area of ​​the internal standard; I iis the peak area of ​​compound i in the sample to be tested.

[0059] Orthogonal partial least squares discriminant analysis is used to decompose X into two types of information related to Y and unrelated to Y, and the difference variables are screened by removing the differences in unrelated orthogonal principal component information;

[0060] After OPLS-DA performs log2 transformation on the original data, it performs central processing and calculates the VIP value of the metabolite variable importance projection. The VIP value represents the influence of the inter-group difference of the corresponding metabolite on the classification and discrimination of each group of samples in the model.

[0061] VIP>1 means significant difference, and the difference fold value is combined to further screen out the differential metabolites between different samples; metabolites with VIP>1 are screened and selected, and then metabolites with fo ld change ≥ 2 and fo ld change ≤ 0.5 are selected. Metabolites with a difference of more than 2 times or less than 0.5 between the control group and the treatment group are considered differential metabolites.

[0062] The random forest algorithm was used to construct the oil grade prediction model, and the selected image and metabolite features were modeled separately.

[0063] The CARS algorithm is a feature variable screening method that combines Monte Carlo sampling with the regression coefficient of the PLS model, imitating the principle of "survival of the fittest" in Darwin's theory. In CARS, each time the points with larger absolute value weights of the regression coefficient in the PLS model are retained as new subsets through adaptive weighted sampling, and the points with smaller weights are removed. Then the PLS model is established based on the new subset. After multiple calculations, the data subset with the smallest root mean square error of interactive verification (RMSECV) of the PLS model is selected as the optimal variable combination. The RF algorithm is an integrated algorithm for classification analysis. It randomly extracts training samples from the original sample set to build multiple models for prediction. By building multiple decision trees and taking majority votes, the category with the majority of results is selected as the final category. Since some invalid interference information in the original data reduces the prediction accuracy of the quantitative model, and high-dimensional data also increases the model complexity and calculation time, the CARS algorithm can combine the attenuation function and adaptive weighted sampling to gradually retain and eliminate feature variables, obtain effective information in the data, reduce redundant information, and finally find the data subset with the smallest RMSECV as the optimal feature combination. The joint application of the CARS algorithm and the RF algorithm can reduce the feature dimension, extract the optimal feature data subset as the model input variable, and improve the detection efficiency of the prediction model.

[0064] like Figure 3As shown in the figure, the random forest (RF) algorithm is used to build the oil grade prediction model, and the selected image and metabolite features are modeled respectively. RF is a supervised machine learning algorithm based on decision trees, which shows excellent performance in classification and regression. Its main feature is that it will not produce overfitting when processing high-dimensional data. In order to find the best score of the RF model, the grid search strategy is used to optimize the parameters of the RF model. GridSearchCV finds the optimal parameter combination in the set parameter range of n_est imators (500, 600, 700, 800), max_depth (None, 5, 10, 15, 20), min_sample_split (2, 5, 10), min_sample_leaf (1, 2, 3, 4, 5, 6, 7, 8), max_features (auto, sqrt, log2), bootstrap (True, False) to avoid falling into the local optimal solution. The model performance evaluation indicators are precision, recall and F1 score.

[0065] like Figure 4 As shown in the figure, the differences in image parameters of tobacco leaf samples with different oil content were analyzed. The results showed that oil content had a certain effect on the morphological characteristics of tobacco leaves. The width, leaf area and roundness of M4 tobacco leaf samples were significantly lower than those of M1, M2 and M3 samples. In terms of color characteristics, M4 leaves were significantly smaller than M1, M2 and M3 samples in RGB color space (redness value and greenness value), Lab color space (brightness value and yellow-blue channel value) and HSV color space (brightness and grayscale value).

[0066] Figure 5The top 20 metabolites with the highest difference in the comparison of different oil categories are shown. Compared with M1, 156 differential metabolites were screened in M2 samples, of which 116 metabolites were significantly upregulated, including methacrylamide (6.12 times), 2,3-dimethylquinolone (6.03 times) and curcumin (6.00 times), and 40 metabolites were significantly downregulated, including cis-3-hexenyl acetate (0.0914 times), trans-3-hexenyl acetate (0.091 times) and limonene (0.093 times). The differential metabolites were mainly concentrated in terpenes and heterocyclic compounds. Compared with M1, M3 screened 120 differential metabolites, of which 71 were upregulated, and the largest difference was ester valeric acid geranyl (9.25 times), and 49 were downregulated, and the largest downregulated was alcohol 2-cyclopentylethanol (0.047 times). Compared with M1, M4 screened out 183 differential metabolites, of which 106 were upregulated. The most significantly upregulated one was the same as that in M3, which was the ester substance geranyl valerate (8.77 times). There were 77 downregulated metabolites, and the ester substance isobutyl benzoate had the largest downregulation (0.027 times).

[0067] like Figure 6 As shown, KEGG pathway enrichment analysis was performed on the differential metabolites of different treatment combinations (M2 / M1, M2 / M1 and M4 / M1), and it was found that compared with M1, 19 differential metabolites in M2 were significantly enriched in the phenylalanine metabolism and diterpene biosynthesis pathways (P<0.05), 11 differential metabolites in M3 were significantly enriched in the diterpene biosynthesis and linolenic acid metabolism pathways (P<0.05), and 15 differential metabolites in M4 were significantly enriched in the diterpene biosynthesis pathway (P<0.05).

[0068] Depend on Figure 7 It can be seen that, compared with the flavor of M1, in the flavor of M2, fruitiness, sweet and green are more prominent, while flora and woody are generally prominent, and rose, waxy, spicy, citrus and fresh are weaker; compared with the flavor of M1, in the flavor of M3, fruitiness, green and flora l are more prominent, while sweet, waxy, woody and rose are generally prominent, and spicy, citrus and herbal are weaker; compared with M1, in the flavor of M4, fruitiness, green and sweet are more prominent, while flora l and woody are generally prominent, and rose, spicy, waxy, citrus and herbal are weaker.

[0069] like Figure 8As shown in Figure A, the CARS algorithm is used to select image features. The number of Monte Carlo sampling is set to 50. It is found that when the iteration reaches 28 times, RMSECV reaches the minimum value of 0.0078, and 8 key image features are extracted, including 5 color features (R_mean, G_mean, B_mean, V_mean and Gray_mean) and 3 texture features (Gray_dissimilarity, R_dissimilarity and G_dissimilarity). Figure 7 As shown in the figure, the CARS method was used to screen 801 volatile metabolites detected by GC-MS. When the sampling run number was 33 times, the RMSECV reached the minimum value of 0.1374. At this time, 16 optimal metabolite feature combinations were selected, mainly including terpenes such as (+)-epi-bicyclosesquiphellandrene, ocimene and curcuminone, alcohols such as 1,2-benzenedimethanol, and aldehydes such as (E,E)-2,4-undecenal.

[0070] like Fig. 9 As shown in the figure, the Pearson correlation coefficient was used to calculate the correlation between 8 image features and 16 metabolite features. In terms of image features, it was found that there was a significant positive correlation between the 5 color features (P<0.05), with correlation coefficients ranging from 0.82 to 0.99. There was also a significant positive correlation between the 3 texture features, but the correlation between color and texture features was low, with correlation coefficients below 0.43. In terms of metabolite features, the four metabolites (+)-episo-bicyclosesquiphellandrene, myrcene, butyl dodecalactone, and 2,6-dimethyl-2-trans-6-octadiene were positively correlated with each other, but were negatively correlated with most other metabolites, among which butyl dodecalactone was significantly negatively correlated with ethyl 2-nonanoate, 7-methyl-3-methylene-6-octenal, and 2-methylbutyric acid isobutyl ester (P<0.05).

[0071] like Fig.10As shown in the figure, the importance of image and metabolite features to the high oil content quality of tobacco leaves was evaluated based on the performance of Mean Decrease Accuracy in the RF algorithm. The results showed that among the eight image features, color information could better reflect the oil quality than texture information, and the Mean Decrease Accuracy values ​​of G_mean and Gray_d issimil ity were the largest, 13.75 and 3.61 respectively, which were the top rankings among the two-dimensional features of color and texture. Among the 16 volatile metabolic features, the Mean Decrease Accuracy values ​​of two ester substances, heptanoic acid methyl ester (Heptano ic acid) and .de lta.-Dodeca lactone (.de lta.-Dodeca lactone), were 17.01 and 14.63 respectively, which contributed more to the accumulation of oil in tobacco leaves.

[0072] like Fig.11 As shown in the figure, the 8 image features and 16 metabolite features selected by CARS were used as input variables, and the data set was split according to a ratio of 7:3. It was found that compared with the default parameters, the use of grid search for the optimal parameters increased the prediction accuracy of the RF model by an average of 0.0967. Finally, the RF model constructed by metabolite features had a prediction accuracy of 87.5% for the M1 oil label, 100% for M2, M3 and M4, and an average model accuracy of 96.88%. However, the RF model established by image features had a poor prediction effect on oil content, with a prediction accuracy of 84% for the M1 grade, and some samples were misidentified as M2. The prediction accuracy of M2, M3 and M4 was all below 80%, with an average accuracy of 69.25%. The ROC curve (Receiver Operating Characteristic curve) and AUC (Area Under the Curve) were used to evaluate the model performance, and it was found that the model performed better in the task of predicting oil content using volatile metabolites. The AUC value of the metabolite feature test set was higher than that of the image feature ( Fig.11 C,D).

[0073] By collecting images of tobacco samples with different oil levels, and using volatile metabolomics technology to analyze the metabolite composition and abundance in tobacco leaves, the quality characteristics of high-oil tobacco leaves were deeply explored. In terms of image processing, the U-Net network was used to segment the background of tobacco leaf images, which is different from traditional image segmentation methods such as threshold segmentation. The threshold segmentation method is actually a point statistical clustering segmentation method, which distinguishes the target image and the background by counting the size of the color attribute value of each pixel. Although this method is simple and has a small amount of calculation, it is only applicable to images with good pixel attribute consistency. In order to make full use of the image features of tobacco leaves themselves and effectively expand the feature parameters of tobacco leaf image recognition, the image segmentation method introduced in this study enables the segmented images to better retain the original feature information of the image, laying the foundation for further using images to achieve quantitative evaluation of tobacco leaves. In terms of metabolite accumulation, compared with low-oil (M1 and M2) tobacco leaves, the content of ethers, heterocyclic compounds, terpenes and halogenated hydrocarbons in high-oil tobacco leaves has increased significantly, and the flavor of tobacco leaves is more prominent in fresh, green and sweet.

[0074] For high-dimensional images and metabolite features, this study used the CARS algorithm to perform feature optimization on the data in the training set, and extracted key features based on the changes in the RMSECV value during the Monte Carlo iteration process. A lower RMSECV value means that the prediction error of the model is smaller and the prediction performance of the established model is better. RMSECV is an indicator to measure the prediction accuracy of the model. It is the average value of the root mean square error (RMSE) of each round calculated during the cross-validation process, which shows the average performance of the model on different data subsets. In addition, the RMSE is above the scale of the original data, and the information is more intuitive. This study used the CARS method for feature optimization. In the image feature selection, RMSECV showed a trend of first increasing and then decreasing, while the RMSECV change trend in the metabolite feature was the opposite. The error value first decreased, indicating that some useless metabolite information was eliminated in the early stage of sampling. The error value then increased, indicating that as the number of sampling times increased, the number of variables gradually decreased, resulting in the elimination of some effective features, reducing the accuracy of the model. The model error change trend of selecting characteristic wavelengths of eggs using the CARS algorithm is consistent, indicating that with the increase in the number of sampling times, the number of variables involved in the operation gradually decreases, and RMSECV will obtain different results. When RMSECV reaches the minimum value, the data subset obtained by sampling at this time is the optimal feature combination. This study selected 8 features from 54 image variables, but the model did not select morphological indicators as input variables, indicating that morphological features may have little effect on oil quality. In addition, CARS selected 16 features from 801 volatile metabolites, mainly terpenes (5), esters (4) and aldehydes (4), which indicates that terpenoid metabolites may be closely related to the oil quality of tobacco leaves.

[0075] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.

Claims

1. Characterization method of tobacco leaf oil based on image processing and volatile metabolite analysis, It is characterized in that The following steps are involved: S100: Place tobacco leaves in a still box and take images. Use image processing technology to extract information such as morphology, color, and texture in the images. Use GC-MS volatile metabolomics to analyze the metabolite components of tobacco leaves of four oil categories, analyze the metabolite differences of tobacco leaves of different oil grades, and use feature selection methods to mine key images and metabolite features that can characterize the oil quality of tobacco leaves. S2 00: The U-Net convolutional network is used to segment the background of the collected tobacco leaf image. The U-Net network structure consists of two parts. The first part is the feature extraction part on the left, and the second part is the upsampling part on the right. S300: Acquire tobacco leaf image feature information, use the Aiseed scanning system to extract tobacco leaf image information, and extract image features, wherein the image features include morphological feature values, color feature values, and texture feature values.

2. The method for characterizing tobacco leaf oil based on image processing and volatile metabolite analysis according to claim 1, It is characterized in that S110: weighing a tobacco leaf sample, grinding it and placing it in a headspace bottle to obtain a powder sample; S120: adding a saturated NaCl solution and an internal standard solution to the powder sample to obtain a mixed solution at a constant temperature of 60° C.; S130: The mixed solution is subjected to headspace extraction using a fully automatic headspace solid phase microextraction device for GC-MS separation and identification; S140: Establish a database, select 1 quantitative ion and 2-3 qualitative ions for each compound and perform precise scanning. Each group of detected ions is detected in the order of peaks and in different time periods. If the detected retention time is consistent with the standard reference, and the selected ions appear in the sample mass spectrum after background subtraction, it is determined to be the substance, and the quantitative ion is selected for integration and correction to enhance quantitative accuracy. At the same time, based on the original data, the missing values ​​are first filled using the k nearest neighbor algorithm, and then the QC sample CV value is calculated, and the substances with CV values ​​less than 0.5 are retained to obtain the final data. In the extraction test, 12 biological replicates were performed for each grade of oil sample.

3. The method for characterizing tobacco leaf oil based on image processing and volatile metabolite analysis according to claim 1, It is characterized in that MassHunter software was used to process the mass spectrometry files and perform integration and correction. choose An isotope internal standard is used to calculate the relative content of VOCs in the sample according to the following formula; The formula for calculating the relative content of compounds in solid samples is: X i is the content of compound i in the sample to be tested (μg / g); V s is the volume of internal standard added (μL); C s is the concentration of the internal standard (μg / mL); M is the amount of the sample to be tested (g); I s is the peak area of ​​the internal standard; I i is the peak area of ​​compound i in the sample to be tested; The formula for calculating the relative content of compounds in liquid samples is: X i is the content of compound i in the sample to be tested (μg / mL); V s is the volume of internal standard added (μL); C s is the concentration of the internal standard (μg / mL); V is the volume of the sample to be tested (mL); I s is the peak area of ​​the internal standard; I i is the peak area of ​​compound i in the sample to be tested.

4. The method for characterizing tobacco leaf oil based on image processing and volatile metabolite analysis according to claim 1, It is characterized in that Orthogonal partial least squares discriminant analysis is used to decompose X into two types of information related to Y and unrelated to Y, and the difference variables are screened by removing the differences in unrelated orthogonal principal component information; After the original data were log2-transformed, OPLS-DA was centered to calculate the VIP value of the metabolite variable importance projection. The VIP value represents the influence of the inter-group difference of the corresponding metabolite on the classification and discrimination of each group of samples in the model. VIP>1 indicates significant difference, and the difference fold value is combined to further screen out the differential metabolites between different samples; metabolites with VIP>1 are selected, and then metabolites with fold change≥2 and fold change≤0.5 are selected. Metabolites with a difference of more than 2 times or less than 0.5 between the control group and the treatment group are considered differential metabolites.

5. The method for characterizing tobacco leaf oil based on image processing and volatile metabolite analysis according to claim 1, It is characterized in that The random forest algorithm was used to construct the oil grade prediction model, and the selected image and metabolite features were modeled separately.

Citation Information

Cited By

  • Dendrobium officinale powder release control method

    CN120507300A

  • Dendrobium officinale powder release control method

    CN120507300B