Jackfruit quality evaluation method based on microminiature hyperspectrum combined with machine learning

Abstract

The invention discloses a jackfruit quality evaluation method based on microminiature hyperspectrum combined with machine learning, and the method comprises the steps: obtaining hyperspectral signals of jackfruits with different maturity degrees under the wave band of 713-920 nm, determining physicochemical indexes according to a traditional method, preprocessing the collected spectral signals through the methods of multivariate scatter correction, standard normal variable transformation and the like, and calculating the quality of the jackfruits with different maturity degrees. A jack fruit quality evaluation model is constructed by combining machine learning methods such as partial least squares, a support vector machine and a one-dimensional convolutional neural network on the basis of characteristic wave band screening algorithms such as a continuous projection algorithm and competitive adaptive reweighted sampling, and effective prediction of the jack fruit storage quality is realized by only using 10 characteristic wave band parameters. According to the method, the maturity and key quality indexes of the jackfruit pulp can be nondestructively detected only by collecting the microminiature hyperspectral signals of the jackfruit pulp, and in the links of planting, storing, processing and selling the jackfruit, the technology can remarkably improve the efficiency and accuracy of quality monitoring and support on-site rapid detection.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of non-destructive testing and spectral analysis of agricultural product quality, specifically involving a method for evaluating the quality of jackfruit based on micro-scale hyperspectral imaging combined with machine learning, which is particularly suitable for rapid detection and evaluation of the quality of jackfruit at different maturity levels. Background Technology

[0002] The quality evaluation of jackfruit involves multiple dimensions, including appearance, internal quality, and storage characteristics. Consumers' quality requirements for jackfruit have expanded beyond simple appearance and freshness to include taste, flavor, and nutritional components. Meanwhile, the industry needs to achieve grading and sorting, storage control, and processing optimization through quality testing. Traditional jackfruit quality testing methods (such as refractometer method for soluble solids, hardness tester method for hardness, and titration method for acidity) can accurately obtain physicochemical indicators, but they are cumbersome, time-consuming, energy-intensive, and can damage samples, making them difficult to meet the industry's needs for large-scale, rapid testing. In particular, jackfruit ripens within a concentrated period and is not resistant to storage, easily suffering mechanical damage and quality deterioration during transportation and storage. Traditional methods cannot achieve early, non-destructive testing, resulting in low efficiency in quality control.

[0003] Hyperspectral imaging technology, as a non-destructive testing method integrating spectral analysis and imaging, offers advantages such as speed, accuracy, non-destructiveness, and the ability to simultaneously acquire spatial and spectral information. It has been widely applied in fields such as variety identification of grain crops and quality and maturity testing of fruits and vegetables. This technology captures the optical response of samples across multiple wavelengths and combines it with chemometric methods (such as partial least squares regression, support vector machines, and deep learning models) to achieve qualitative and quantitative analysis of crop quality indicators. For example, hyperspectral technology has been successfully used to detect sugar content in watermelons and cantaloupes and surface defects in kiwifruit, achieving a high level of prediction accuracy.

[0004] However, the application of hyperspectral technology in jackfruit quality evaluation still has significant gaps. On the one hand, jackfruit pulp has a complex composition, and its quality indicators exhibit unique changes during ripening, differing from other fruits and vegetables, making it difficult to directly apply existing spectral models. On the other hand, the optical signal of jackfruit is easily affected by factors such as variety and storage conditions, resulting in significant spectral band overlap and interference issues, requiring the integration of characteristic band screening algorithms to improve model accuracy and generalization ability. Furthermore, existing research largely focuses on large-scale hyperspectral equipment, while miniature hyperspectral equipment is more suitable for on-site testing due to its portability, but related technical solutions are not yet mature.

[0005] Therefore, developing a quality detection method for jackfruit based on micro-miniature hyperspectral technology combined with chemometrics can solve the pain points of traditional detection methods, fill the gap in the application of hyperspectral technology in jackfruit, and has important scientific value and application prospects for improving the quality control level of the jackfruit industry and promoting the efficient development of the industry. Summary of the Invention

[0006] The purpose of this invention is to provide a method for evaluating the quality of jackfruit based on micro-hyperspectroscopy combined with machine learning. By using micro-hyperspectroscopy, optical signals of jackfruit at different maturity levels are acquired and their quality indicators (chromaticity L*, a*, b*, C*, H*) are measured.

[0007] Soluble solids, titratable acidity, etc. are preprocessed using multivariate scattering correction (MSC) and standard normalization (SNV). Feature band filtering algorithms such as continuous projection algorithm (SPA), competitive adaptive reweighted sampling algorithm (CARS), and non-information variable elimination (UVE) are combined with partial least squares (PLS), support vector machine (SVM), and one-dimensional convolutional neural network (1D-CNN) algorithms to analyze and construct an optical correlation system for jackfruit quality, so as to realize the detection of jackfruit quality by optical characteristics.

[0008] Specifically, the following steps are included:

[0009] S1. Fresh jackfruit of two varieties, "Xiangmi No. 1" and "Malaysia No. 1," were selected. The pulp was removed, and the pulp with excellent quality, uniform size, no significant differences, no pests or diseases, no mechanical damage, and no other defects was selected. The samples were divided into 5 groups according to different maturity levels and stored in a cold storage at 4℃ for 0, 2, 4, 6, and 8 days, respectively, for a total of 360 jackfruit samples = 2 fruit varieties × 5 storage periods × 6 replicates × 6 jackfruit pulp samples / group;

[0010] S2. Using a Monarch Pro portable multispectral camera (Unico Corporation, Israel) and an XD 301 fiber optic cold light source, hyperspectral images of a calibration white board, calibration black board, and jackfruit sample were acquired sequentially in 10 wavelength bands (713, 736, 759, 782, 805, 828, 851, 874, 897, 920 nm). The reflectance of the samples in the 10 wavelength bands was calculated using the following formula:

[0011]

[0012] In the formula , , , These represent the sample reflectance, the light intensity reflected from the white plate, the light intensity reflected from the dark plate, and the light intensity reflected from the sample, respectively.

[0013] S3. Using a CR-400 portable colorimeter (Minolta, Japan), the lightness (L*), red-green value (a*), yellow-blue value (b*), saturation (C*), and hue (H*) of the fruit pulp were measured. Each sample was measured three times, and the average value was taken. A PAL-1 handheld saccharimeter (Aitaka, Japan) was used to directly measure the fruit pulp juice, and the results were expressed as Brix (°Brix). The instrument was calibrated with distilled water before measurement. A pH meter (Aitaka, Japan) was used for measurement. The above jackfruit juice was dripped into the detection port to measure the titratable acid content, and the results were expressed as a percentage (%).

[0014] S4. SPSS 27.0 (IBM SPSS, USA) was used to perform single-factor and correlation analyses on the physicochemical indicators of jackfruit, such as color, soluble solids, and titratable acidity. Matlab 2023a (MathWorks, USA) was used to extract the preprocessed hyperspectral feature information. Partial Least Squares (PLS), Support Vector Machine (SVM), and One-Dimensional Convolutional Neural Network (1D-CNN) models were constructed using Python 3.8, and the optimal model was selected by using indicators such as calibration set determination coefficient (Rc²), validation set determination coefficient (Rv²), and root mean square error (RMSE).

[0015] The advantages of this invention are:

[0016] (1) The variation law of jackfruit quality (color, soluble solids, titratable acidity, etc.) and optical properties was explored, and a high-throughput optical database of jackfruit was constructed;

[0017] (2) A micro-scale hyperspectral predictive model for jackfruit quality was developed, and the characteristic band screening method was used to digitally evaluate the quality of jackfruit.

[0018] (3) A method for rapidly predicting the quality of jackfruit has been developed, which simplifies the detection steps for jackfruit quality. Compared with traditional jackfruit quality determination, this invention saves time, reduces labor, and significantly reduces the testing cost.

[0019] (4) By rapidly characterizing the optical properties of jackfruit, the quality of jackfruit can be predicted quickly, enabling food companies to achieve rapid, efficient and stable product quality control. Attached Figure Description

[0020] Figure 1 : Flowchart of the jackfruit quality testing method of this invention;

[0021] Figure 2 Changes in the apparent state of jackfruit during different storage periods;

[0022] Figure 3Color changes of two types of jackfruit during different storage periods (a:L) * Value; b: a * Value; c: b * Value; d: C * value;

[0023] e:H * value);

[0024] Figure 4 Changes in soluble solids and titratable acidity of two types of jackfruit under different storage periods (a: SSC; b: TA);

[0025] Figure 5 Changes in reflectance spectra of two types of jackfruit during storage (a: "Xiangmi No. 1"; b: "Malaysia No. 1").

[0026] Figure 6 Correlation analysis between jackfruit quality indicators and spectral data;

[0027] Figure 7 Confusion matrices of different models based on jackfruit variety classification (a: PLS; b: SVM; c: 1D-CNN);

[0028] Figure 8 Classification confusion matrix of jackfruit under storage time (a: PLS; b: SVM; c: 1D-CNN); Detailed Implementation

[0029] The specific implementation method of jackfruit quality evaluation based on micro-scale hyperspectral imaging combined with machine learning is as follows:

[0030] I. Experimental Materials and Methods

[0031] 1. Sample preparation:

[0032] The experimental samples were fresh jackfruit (two varieties, "Xiangmi No. 1" and "Malaysia No. 1") purchased from a local fruit wholesale market in Nanjing, Jiangsu Province. Jackfruit without any obvious defects, diseases, or unfresh color were selected. The jackfruit pulp was extracted, and healthy, undamaged pulp was stored in a 4℃ cold storage for 0, 2, 4, 6, and 8 days respectively. A total of 360 jackfruit samples were obtained: 2 fruit varieties × 5 storage periods × 6 replicates × 6 jackfruit pulp samples per replicate.

[0033] 2. Hyperspectral image acquisition and correction:

[0034] Using a Monarch Pro™ portable multispectral camera (Unico Corporation, Israel) and an xd 301 fiber optic cold light source (Shanghai Wence Optoelectronic Technology Co., Ltd.), hyperspectral images of a calibration white board, calibration black board, and jackfruit sample were acquired sequentially in 10 wavelength bands (713, 736, 759, 782, 805, 828, 851, 874, 897, and 920 nm). The reflectance of the samples in the 10 wavelength bands was then calculated using the following formula.

[0035]

[0036] In the formula , , , These represent the sample reflectance, the light intensity reflected from the white plate, the light intensity reflected from the dark plate, and the light intensity reflected from the sample, respectively.

[0037] 3. Determination of physicochemical indicators:

[0038] (1) Chromaticity

[0039] The jackfruit fruit samples collected via hyperspectral imaging were immediately subjected to colorimetric determination. The color of the jackfruit pulp was measured using a CR-400 colorimeter (Minolta, Japan), and the results were compared with CIE 1976 L. * a * b * Lightness value in color space (L) * ), red-green value (a * ), yellow-blue value (b) * ), saturation (C) * ), Hue (H) * )express.

[0040] (2) Soluble solids

[0041] The SSC content of jackfruit was determined using a PAL-1 handheld saccharimeter (ATRO Corporation, Japan), with a measurement range of 0-30 °Brix. Before measurement, the instrument was calibrated with distilled water and the detection prism was cleaned. The jackfruit pulp was then crushed using a crusher to extract the juice for testing. The readings were allowed to stabilize and recorded. The results were expressed as Brix (°Brix).

[0042] (3) Titratable acid

[0043] The titratable acid content was determined using a PAL-Easy ACID F5 pH meter (ATRO Corporation, Japan). The above-mentioned jackfruit juice was dripped into the detection port to determine the titratable acid content, and the result was expressed as a percentage (%).

[0044] II. Spectral Preprocessing and Model Construction

[0045] 1. Data Preprocessing

[0046] Matlab 2023a was used to perform multivariate scattering correction (MSC), standard normal variable transformation (SNV), and normalization on the hyperspectral data to eliminate the effects of noise and baseline drift.

[0047] 2. Feature band selection

[0048] Continuous projection algorithm (SPA), competitive adaptive reweighted sampling (CARS), and non-information variable elimination (UVE) were used to select characteristic bands. The results showed that the 713–920 nm band had the highest correlation with indicators such as SSC and chromaticity, and 10 characteristic bands were ultimately retained for modeling.

[0049] 3. Model Building and Validation

[0050] Partial Least Squares (PLS), Support Vector Machine (SVM), and One-Dimensional Convolutional Neural Network (1D-CNN) models were constructed using Python 3.8, with a randomized partitioning of the calibration and validation sets. The coefficient of determination (R²) of the calibration set was used to determine the validation set. c 2 ), validation set determination coefficient (R²) v 2 The model performance is evaluated using the root mean square error of the calibration set (RMSEC) and the root mean square error of the validation set (RMSEV).

[0051] III. Analysis of Changes in the Quality Characteristics of Jackfruit

[0052] like Figure 2 As shown, the overall color difference of jackfruit at different ripening stages is not significant. Excluding the objective influence of chilling injury, it is difficult to accurately determine the ripeness and internal quality of jackfruit solely based on external images. Therefore, it is necessary to accurately determine the quality of jackfruit based on the SSC content, TA value, color, and spectral differences at different storage periods.

[0053] 1. Color change

[0054] from Figure 3It can be seen that under storage conditions of 4℃, there are significant differences in the color indicators of "Xiangmi No. 1" and "Malaysia No. 1" jackfruit in terms of brightness, red-green, yellow-blue, saturation and hue, and these differences show a specific pattern as they change with the storage period (0-8 days). Specifically, in terms of brightness, "Xiangmi No. 1" ranges from 60 to 75, with a higher brightness than "Malaysia No. 1" (45-65). The brightness of both decreases with increasing storage time. In terms of red-green values, "Xiangmi No. 1" ranges from 15 to 25, with a deeper orange hue (red leaning towards yellow) than "Malaysia No. 1" (5-15), and both decrease with storage time. In terms of yellow-blue values, "Xiangmi No. 1" ranges from 50 to 55, with a more pronounced yellow hue, while "Malaysia No. 1" ranges from 35 to 50. Both decrease with storage time. In terms of saturation, "Xiangmi No. 1" ranges from 50 to 60, higher than "Malaysia No. 1" (40-50), and both decrease with storage time. In terms of hue, "Xiangmi No. 1" ranges from 65 to 75, while "Malaysia No. 1" ranges from 75 to 80, leaning more towards yellow-green, and the hue increases with storage time.

[0055] 2. Changes in soluble solids

[0056] like Figure 4 As shown, the overall trend of both "Xiangmi No. 1" and "Malaysia No. 1" jackfruit was an initial increase followed by a decrease. There was no significant difference at 0 days of storage (P < 0.05), but significant differences were observed at 2, 4, and 8 days of storage (P < 0.05), with the variation range between 15 and 30. The SSC content reached its highest level at 6 days of storage. Based on the analysis of significant differences at different storage periods, "Xiangmi No. 1" showed a significant difference between 0 and 2 days of storage (P < 0.05), while "Malaysia No. 1" showed significant differences between 0, 2, and 4 days of storage (P < 0.05).

[0057] 3. Titratable acid change

[0058] like Figure 4 As shown, both "Xiangmi No. 1" and "Malaysia No. 1" jackfruit exhibited an overall trend of first increasing and then decreasing, which is opposite to the trend of SSC content changes. There was no significant difference between the two jackfruit varieties at 0 days of storage (P < 0.05), but significant differences were observed at 2 and 4 days of storage (P < 0.05). Later in the storage period, the differences tended to disappear (P < 0.05), with the variation range being 0.12 to 0.18. The TA content reached its lowest point at 2 days of storage. Analysis of significant differences across different storage periods revealed a significant difference at 2 days of storage compared to other storage periods (P < 0.05).

[0059] 4. Optical changes

[0060] like Figure 5 As shown, the reflectance of both "Xiangmi No. 1" and "Malaysia No. 1" jackfruit samples exhibited a trend of first increasing and then decreasing within the wavelength range of 720–920 nm. The reflectance reached a relative peak in the 780–800 nm band before declining. The reflectance variation of the "Malaysia No. 1" jackfruit sample was more complex, with curves intersecting at different storage periods. There were no significant differences in reflectance between the two jackfruit varieties at storage periods of 6 days and 8 days, with ranges of 75–85 and 65–75, respectively. Significant differences were observed at storage periods of 0 days, 2 days, and 4 days, with ranges of 75–80 and 60–70, 70–80 and 60–70, and 80%–90% and 70–85, respectively.

[0061] IV. Correlation Analysis between Jackfruit Quality Indicators and Spectra

[0062] from Figure 6 As can be seen, the correlations between various quality indicators of jackfruit and the spectral reflectance in the 713–920 nm wavelength band exhibit different characteristics. Brightness value (L*) shows a weak positive correlation (0.07–0.26), with the correlation slightly strengthening with increasing wavelength but remaining generally insignificant. Red-green value (a*) shows a positive correlation (0.16–0.39), with the correlation increasing with increasing wavelength, and a closer association with longer wavelength reflectance, possibly related to the reflectance characteristics of the red spectral region. Yellow-blue value (b*) shows a weak positive correlation (0.05–0.25), with small fluctuations and a relatively weak correlation with wavelength. Saturation (C*) shows a moderate positive correlation (0.09–0.29), with the correlation increasing with wavelength. The correlation is slightly stronger; hue (H*) shows a negative correlation (-0.14~-0.39), which strengthens with increasing wavelength and shows a significant inverse relationship with long-wavelength reflectance, possibly related to the absorption differences of carotenoids and chlorophyll at different wavelengths; soluble solids (SSC) show a strong positive correlation (0.43~0.50), with a tight overall correlation, indicating the feasibility of hyperspectral technology for detecting SSC; titratable acid (TA) shows a weak negative correlation (-0.09~-0.15), with no significant correlation, and the prediction accuracy needs further optimization.

[0063] V. Comparison of Model Prediction Results

[0064] The model prediction results show that different algorithms have significant differences in their ability to identify and regress the quality of jackfruit, and are significantly affected by the type of indicator and the preprocessing method.

[0065] In variety identification, the 1D-CNN model performed best. As shown in Table 1, its accuracy on both the calibration and validation sets reached 98.89%, and its precision, recall, and F1 score were completely consistent, indicating that the model has extremely strong stability in classifying "Xiangmi No. 1" and "Malaysia No. 1". Combined with... Figure 7 The confusion matrix shows that the model made zero errors in classifying "Malaysia No. 1" and only one error in classifying "Xiangmi No. 1," significantly outperforming SVM and PLS. Although SVM performed close to 1D-CNN on the calibration set, its accuracy on the validation set dropped significantly, presumably due to insufficient generalization ability caused by uneven sample distribution. PLS, as a linear model, struggled to capture the complex nonlinear differences between varieties, resulting in a high misclassification rate.

[0066] Table 1 Comparison of model discriminant performance based on different jackfruit varieties

[0067]

[0068] For determining storage time, 1D-CNN also demonstrates advantages. Table 2 shows that its validation set accuracy reaches 88.89%, indicating good performance of the 1D-CNN model. It performs excellently on the calibration set, although the validation set accuracy is slightly lower than the calibration set accuracy, it still remains at a high level. This indicates that the model's judgment ability is significantly better than other models, and its generalization ability is strong but still needs improvement. Figure 8 The 1D-CNN model shown has a very high proportion of diagonal elements, with only a small number of misclassifications. In contrast, SVM and PLS exhibit severe misclassifications. Therefore, the 1D-CNN model demonstrates the best discriminative ability.

[0069] Table 2 Comparison of model discriminant performance based on different storage times of jackfruit

[0070]

[0071] In the regression model, the predictability of different indicators varied significantly. Table 3 shows that hue (H*) had the best predictive performance. When 1D-CNN was combined with SNV or MSC preprocessing, the residual prediction (RPD) reached 2.27 (>2.0), the coefficient of determination (Rc²) on the calibration set was 0.92, and the Rv² on the validation set was 0.81, indicating that the model could stably capture the hue shift caused by carotenoid degradation and chlorophyll changes. The predictive performance of color indicators such as brightness (L*) and red-green value (a*) was moderate (RPD 1.33-1.89). Among them, 1D-CNN, after SNV preprocessing, showed better prediction performance for L* (RPD 1.89) than SVM (1.78) and PLS (1.54), demonstrating the analytical advantage of nonlinear models for complex associations.

[0072] Among chemical indicators, soluble solids (SSC) showed better predictive performance than titratable acid (TA). When SVM was combined with SNV pretreatment, the RPD of SSC reached 1.43, while the optimal RPD of TA was only 1.14. This is consistent with... Figure 6The correlation analysis results were consistent—SSC showed a strong positive correlation with reflectivity in the 713-920nm band (r=0.43-0.50), while TA showed a weaker correlation (r=-0.09 to -0.15). The 1D-CNN prediction for SSC fluctuated significantly; the RPD was only 1.15 in the original spectrum, but after SNV processing, the calibration set Rc² increased sharply to 0.94, indicating the crucial role of preprocessing in eliminating noise and enhancing the effective signal.

[0073] Table 3 Comparison of Jackfruit Quality Prediction Model Results

[0074]

[0075]

[0076] Note: Rc2: Coefficient of determination of calibration set (training set); Rv2: Coefficient of determination of validation set (validation set); RMSEC: Root mean square error of calibration set; RMSEV: Root mean square error of validation set; MSC: Multivariate scattering correction; SNV: Standard normal variable; RPD: Remaining forecast residuals.

[0077] In summary, this study utilizes miniature hyperspectral technology, combined with feature band selection and machine learning algorithms, to evaluate the quality of jackfruit. Practice shows that building a model by selecting key bands reduces model complexity while improving prediction accuracy. The 1D-CNN model, in particular, demonstrates outstanding performance in variety classification, storage time determination, and prediction of some quality indicators, providing an effective solution for rapid and non-destructive testing of jackfruit quality. This research supports the application of miniature hyperspectral technology in all stages of the jackfruit industry and has practical significance for improving the efficiency of quality control.

Claims

1. A method for evaluating the quality of jackfruit based on micro-hyperspectral imaging combined with machine learning, comprising the following steps: S1.Select fresh jackfruits of two varieties, "Xiangmi No. 1" and "Malaysia No. 1", peel the flesh, and select the flesh with excellent quality, uniform size, no disease and insect pests, no mechanical damage, and no other defects; S2.Use Monarch Pro portable multispectral camera and xd 301 type optical fiber cold light source to obtain hyperspectral image information of the calibration white board, calibration black board and jackfruit sample under 10 wavebands in turn; S3.Use CR-400 portable color difference meter to measure the lightness, red-green value, yellow-blue value, saturation and hue of the flesh, measure each sample 3 times and take the average value; S4.Use SPSS 27.0 to perform single factor analysis and correlation analysis on the color, soluble solids and titratable acidity of jackfruit.

2. The method for quality evaluation of jackfruit based on micro-hyperspectral imaging combined with machine learning according to claim 1, characterized in that, In S1, the samples are divided into 5 groups according to different maturity, and stored in a 4℃ refrigerator for 0, 2, 4, 6 and 8 days, respectively. A total of 360 jackfruit samples = 2 fruit varieties × 5 storage periods × 6 repeated experiments × 6 jackfruit flesh per group.

3. The method for quality evaluation of jackfruit based on micro-hyperspectral imaging combined with machine learning according to claim 1, characterized in that, In S3, the PAL-1 handheld sugar meter is used to directly measure the juice of the flesh, and the result is expressed in Brix. The instrument is calibrated with distilled water before measurement. The PAL-Easy ACID F5 type acidimeter is used for measurement. The juice of the jackfruit is taken and dropped into the detection port to determine the titratable acid content.

4. The method for quality evaluation of jackfruit based on micro-hyperspectral imaging combined with machine learning according to claim 1, characterized in that, In S4, Matlab 2023a is used to extract the preprocessed hyperspectral feature information; Python 3.8 is used to build partial least squares, support vector machines and one-dimensional convolutional neural network models, respectively, and the optimal model is selected by the determination coefficient of the calibration set, the determination coefficient of the validation set and the root mean square error.

5. The method for quality evaluation of jackfruit based on micro-hyperspectral imaging combined with machine learning according to claim 1, characterized in that, The reflectivity of the samples under 10 wavebands is calculated by the following formula: wherein , , , represent the reflectivity of the sample, the white plate reflected light intensity, the dark plate reflected light intensity, the sample reflected light intensity, respectively.

Images