Non-destructive prediction of anthocyanin concentration in whole eggplant peel using hyperspectral imaging

Plant material

A total of 20 eggplant varieties bred by the Chinese Academy of Tropical Agricultural Sciences, South Subtropical Crop Research Institute were selected. These 20 eggplant varieties encompass a range of colors including white, green, light-purple, green-purple, and dark-purple. In total, 277 eggplant fruits were collected, consisting of 10 samples of white color, 28 samples of green color, 111 samples of light-purple color, 17 samples of green-purple color, and 111 samples of dark-purple color (Table S1).

Hyperspectral image acquisition

The imaging system comprised of a SOC70VP hyperspectral camera and a lamp holder. The hyperspectral camera covered wavelengths ranging from 400 to 1,000 nm with an approximate resolution of 0.6 nm with 128 pixels (channels) in the wavelength dimension. The lamp holder accommodated two Philips halogen lamps with a power rating of 500 Watts at 220 volts. The eggplant and spectralon were positioned beneath the hyperspectral camera to allow reflection of the light emitted by the halogen lamps.

Following hyperspectral imaging, one mm uniformly thick peel was collected using a sharp knife immediately, frozen using liquid nitrogen, and stored at −80 °C for subsequent anthocyanin content analysis. Reflectance values were calculated using the SRAnal 710 software. ROI extraction and the calculation of the average reflectance were performed using ENVI 5.3.

Anthocyanin extraction

The identification and quantification of anthocyanins in eggplant peel were conducted using High Performance Liquid Chromatography (HPLC) methods. Six types of anthocyanins were detected, namely delphinidin, cyanidin, petunidin, pelargonidin, paeonidin, and malvidin. The extraction and hydrolysis methods for anthocyanins were adjusted based on the procedure outlined in HPLC(2014). The extraction process involved a solution consisting of anhydrous ethanol, water, and hydrochloric acid in a ratio of 2:1:1. To initiate the extraction, 1.0 g of powder was accurately weighed and transferred into a 10 mL volumetric flask with a stopper. The extractant was then added to the mark, and the mixture was vigorously shaken for 1 min. Ultrasonic extraction was subsequently performed under light-protected conditions for 30 min. For the hydrolysis of anthocyanins to anthocyanidins, the extract obtained from ultrasonic extraction underwent a boiling water bath for 1 h. After cooling, additional extractant was added to bring the total volume to 10 mL. The mixture was thoroughly shaken and allowed to settle. The supernatant was collected and filtered through a 0.22 µm organic membrane.

HPLC analysis of anthocyanins

The HPLC analysis followed the reference method for determining anthocyanidins in plant origin products-High performance liquid chromatography in Deineka & Grigor’ev (2004). The HPLC system utilized for the analysis was the LC-20A equipped with a UV detector. A mobile phase A consisting of formic acid and water in a ratio of 1:9 was used for the HPLC analysis. Meanwhile, mobile phase B consisted of methanol, acetonitrile, water, and formic acid in a ratio of 22.5:22.5:40:10. The analysis employed a gradient elution method with specific time intervals and percentages of mobile phase B as follows: 0–2 min: 7–40% B; 2–11 min: 40–67% B; 11–12 min: 67–100% B; 12–14 min: 100% B; 14–15 min: 100–7% B; 15–20 min: 7% B. Each sample was injected with a volume of 10 µl, and three replicates were performed for each sample.

The standards for delphinidin, cyanidin, petunidin, pelargonidin, peonidin, and malvidin were acquired from Sigma. The total content of anthocyanin was determined by summing the quantities of these six types of anthocyanins.

Hyperspectral data preprocessing

To reduce the influence of factors such as background noise on the average reflectance, pre-processing was performed on the average reflectance data, as shown in Table 1 using Matlab 2020a (The MathWorks Inc., Natick, MA, USA). Following pre-processing, the data were divided into a training dataset and a test dataset at a ratio of 7:3 using the randperm function.

Feature variables extraction

To reduce the number of input variables and improve the model efficiency, the spectral data features were extracted by CARS as described by Li et al. (2009) with the default parameters.

Modeling algorithms

The LS-SVM were implemented using Matlab 2020a with LS-SVMlabv1_8. PLSR were completed by using Matlab 2020a. The flowchart of PLSR analysis is shown in Fig. 1. The following statistical parameters were calculated:

root mean square error of prediction (RMSEP).

root mean square error of calibration (RMSEC). ${\displaystyle \text{RMSE}=\sqrt{\frac{1}{\mathrm{n}}\sum _{\mathrm{i}=1}^{\mathrm{n}}{\left(\text{Measurementi-Predicatedi}\right)}^2}}$

R²_p (coefficient of determination of Prediction).

R²_C (coefficient of determination of Calibration). ${\displaystyle {\mathrm{R}}^2=1-\frac{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\left(\text{Meansurementi}-\text{Predicatedi}\right)}^2}{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\left(\text{Measurementi}-\text{Mean}\left(\text{Meantsurement}\right)\right)}^2}}$

RPDp (Ratio of standard deviation of the validation set to standard error of prediction of Prediction).

RPDc (Ratio of standard deviation of the validation set to standard error of prediction of Calibration). ${\displaystyle \mathrm{RPD}=\frac{\mathrm{SD}}{\text{RMSEP}}.}$

Various eggplant varieties havedifferinglevels of anthocyanin content anddemonstrate distinctaverage reflectance spectra

The results indicate that eggplants of different colors contain varying types of anthocyanins (Figs. 2A and 2B, Table S1). The peels of 138 eggplants contained cyanidin and delphinidin, while the peels of 67 eggplants contained petunidin and delphinidin. Additionally, 21 eggplants contained cyanidin, delphinidin, and petunidin. Notably, none of the 277 eggplants analyzed contained pelargonidin, peonidin, or malvidin. Except for two white eggplants that contained 0.9834 µg/g and 0.6368 µg/g of delphinidin, the remaining white and green eggplants had undetectable anthocyanin content. Cyanidin content ranged from 0 to 9.7430 µg/g, delphinidin content ranged from 0 to 660.177 µg/g and petunidin content ranged from 0 to 17.3905 µg/g. The total anthocyanins content ranged from 0 to 668.049 µg/g.

The average reflectance varied among eggplants of different colors. Both the green and green-purple fruits exhibited an average reflectance model characteristic of green plants, with a peak at 550 nm. However, the green-purple eggplant fruit had lower reflectance compared to the green eggplant fruit due to the presence of anthocyanins in the green-purple variety. White eggplants demonstrated the highest reflectance between 400–700 nm, while dark-purple eggplants had the lowest reflectance. Light-purple and dark-purple eggplants displayed a minimum reflectance around 500 nm (Fig. 2C).

Hyperspectral data preprocessing analysis

In order to establish reliable prediction models, we applied five pretreatment methods to the spectral data and extracted the feature variables using CARS. The results indicated that the average reflectance became more concentrated after pretreatment, compared to the non-preprocessed reflectance (Fig. 3).

To enhance the accuracy and robustness of the diagnostic models, the CARS were utilized to extract feature variables from the pool of 128 variables. In the case of petunidin, only five feature variables were present without any preprocessing. However, after pretreatment, the number of feature variables increased to 128 for SNV, six for AUT, five for NOR, 12 for SG, and four for MC pretreatment, as shown in Table S2 and Fig S1. Similarly, in the case of cyanidin, there were initially only five feature variables without any preprocessing. However, after pretreatment, the number of feature variables increased to 128 for SNV, six for AUT, five for NOR, 12 for SG, and four for MC pretreatment, as indicated in Table S2 and Fig S2. Likewise, for delphinidin, the number of feature variables was initially 5 without any preprocessing. However, following pretreatment, the number increased to 128 for SNV, six for AUT, 24 for NOR, 20 for SG, and five for MC pretreatment, as provided in Table S2 and Fig S3. Finally, in the case of total anthocyanins, the initial count of feature variables without any preprocessing was five. However, after pretreatment, the number increased to 128 for SNV, 25 for AUT, nine for NOR, 32 for SG, and 15 for MC pretreatment, as described in Table S2 and Fig S4.

Modeling and validation of the regression models

In this study, the PLSR and LS-SVM models were utilized to develop estimation models for cyanidin, delphinidin, petunidin, and total anthocyanins. The performance of the PLS regression model on the cyanidin was evaluated. Despite the SNV-PLSR model having the highest R_c² (0.3604) and R²_P (0.3825), the respective RMSEC and RMSEP values were 3.3993 and 3.6949 (Fig. 4). This result indicates a significant disparity between the measured and predicted cyanidin content. Interestingly, the LS-SVM model demonstrated superior predictive performance compared to the PLSR model. However, the perfect predictive outcome of the SNV LS-SVM model may be unreliable due to warnings from Matlab software suggesting potential issues with singularity or improper scaling, thereby resulting in inaccurate predictions. The NOR preprocessing LS-SVM model achieved the best results with with an R_c² value of 9955, and R_p² value of 0.9918. Additionally, the RMSEC and RMSP values were 0.0183 and 0.0275, respectively, The RPDc and and RPDp values were 14.5478 and 10.91075, respectively (Fig. 5).

To further enhance prediction accuracy, feature variables extracted by CARS were used to build the LS-SVM model. Similar caution should be exercised regarding the reliability of the SNV-CARS LS-SVM model. The optimal model was the NOR-CARS preprocessed LS-SVM model, which achieved R_c² and R_p² values of 0.9953 and 0.9880, respectively. Additionally, the RMSEC and RMSEP values were 0.0195 and 0.0303, respectively. Moreover, the RPDc and and RPDp values were 14.5494 and 8.9784, respectively (Fig. 6).

Moving on to the modeling and validation of delphinidin, it was found that the predictions from the PLSR model were unreliable. Applying preprocessing techniques such as SNV, AUT, NOR, SG, and MC did not improve the accuracy rate of the predictions (Fig. 7).

In contrast, for the LS-SVM model, the SNV LS-SVM also exhibited a perfect accuracy rate, but caution must be exercised regarding its reliability. The NOR LS-SVM model yielded the highest accuracy rate, with an R_c² value of 1.000 and R${}_{\mathrm{p}}^2$ value of 0.9967. Moreover, the corresponding values for RMSEC and RMSEP were notably low, measuring 0.0000 and 0.0203, respectively. Additionally, the RPDc and and RPDp values were 31,992.5331 and 17.7166, respectively (Fig. 8).

Furthermore, for the NOR-CARS LS-SVM model, the predicted result achieved an R_c² value of 0.9997 and R${}_{\mathrm{p}}^2$ value of 0.9959 The corresponding values for RMSEC, RMSEP were notably low, measuring 0.0061, 0.0210, respectively and the the RPDc and and RPDp values were 57.9594 and 15.3560, respectively (Fig. 9).

Then, in the modeling and validation of petunidin, the PLSR analysis indicated that all of the predicted outcomes were not satisfactory. The SNV PLSR model displayed the highest level of accuracy, with an R_c² value of 0.6938 and R_p² value of 0.5935. However, the RMSEP and RMSEC values for this model were 2.7460 9 and 2.7413 respectively. Additionally, the RPDc and and RPDp values were 1.9109 and 1.6605, respectively (Fig. 10).

Similarly, the SNV LS-SVM model exhibited a perfect accuracy rate in its predictions, although caution should be exercised regarding their reliability. The NOR LS-SVM model had the highest accuracy rate, with an R_c² value of 0.9979 and R_P² value of 0.9977. Furthermore, the RMSEC and RMSEP values for this model were 0.0141 and 0.0158 respectively. Additionally, the RPDc and and RPDp values were 21.8766 and 21.0938, respectively (Fig. 11).

Additionally, the NOR-CARS LS-SVM model produced an R_c² value of 0.9894. and R_p² value of 0.9889, with corresponding RMSEC and RMSEP values of 0.0339 and 0.0311 respectively. Additionally, the RPDc and and RPDp values were 9.4040 and 9.6236 respectively (Fig. 12).

Lastly,the results of the PLSR analysis indicated that all of the predicted outcomes for total anthocyanin were not satisfactory(Fig. 13). Similarly, the SNV LS-SVM model displayed a high level of accuracy in its predictions. Nevertheless, these results should be interpreted with caution, as they may not be entirely reliable. The NOR LS-SVM model produced the highest accuracy rate, with an R_c² value of 1.0000 and R_p² value of 0.99673. The corresponding RMSEC and RMSEP values were 0.0009 and 0.0161 respectively. Additionally, the RPDc and and RPDp values were 375.334 and 19.2105 respectively (Fig. 14).

Furthermore, the NOR-CARS LS-SVM model yielded promising predicted results. Its R_c² value of 0.999 and R_p² value of 0.997. The corresponding RMSEP and RMSEC values were 0.0036 and 0.0202 respectively. Additionally, the RPDc and and RPDp values were 91.0856 and 18.4801 respectively (Fig. 15).