Yande Liu

Early fungal infection of citrus is one of the common diseases found during the storage period of citrus, and fungus that infects citrus will spread to the entire batch of citrus as the degree of infection deepens, causing enormous economic losses. Therefore, early detection of fungal infection of citrus is fundamental. The purpose of this study is to explore the qualitative identification of early fungal infections in citrus by using Fourier transform near-infrared (FT-NIR) combined with a variety of chemometric methods. First, discrete wavelet transform (DWT) is used to filter the noise of the spectral signal, then combined with a PLS-DA model, that helps discriminate healthy from infected Citrus. Subsequently, four different feature variable selection methods were introduced, Then, the linear discriminant analysis (LDA) and support vector machine (SVM) two classifiers were combined to establish a qualitative model for the degree of fungal infection. The modeling results show that the SVM modeling effect is better than LDA, and the DWT-CARS-SVM based on the RBF kernel function has the best result, the accuracy rates of the training set and test set are 100% and 97%. The results indicate that FT-NIR spectroscopy, combined with chemometric methods, is able to distinguish early fungal infections in citrus.

Citrus is one of the world’s most famous types of fruits, and its planting areas cover much of the world. China is one of the vital origins of citrus, with rich resource types and many advantageous varieties (1). Wogan oranges are a late-maturing hybrid citrus variety. The maturity period of Wogan is from February to March, and it belongs to a class of spring oranges. Its natural fruit time can be ripe from around February to August, with a sweet and sour taste that is popular with many people. Citrus fungal infection is a common disease found during citrus storage, mainly through fruit skin wounds, and it is one of the most harmful citrus diseases. At the same time, citrus infected by fungi can also infect surrounding citrus, leading to cross-infection and resultant enormous economic loss (2–3). With rapid societal developments, people now have higher standards for the safety and taste of fruits and vegetables. Since the fungal infection site appears very similar to healthy skin, it is difficult to observe the infection with the naked eye. Therefore, detecting early fungal infection in citrus is a very challenging task. At present, the detection of fungal infections of fruit on the packaging line is mainly carried out by using the fluorescence effect of ultraviolet light. Although this method is effective, it can affect the physical and mental health of certain workers exposed to ultraviolet light for a long period of time, due to its high energy (2,4). Therefore, it is necessary to explore a nondestructive, rapid, safe, and reliable detection technique for the early identification of citrus fungal infections.

Near-infrared spectroscopy detection technology is very informative, and includes nearly all the characteristic information of X-H hydrogen-containing groups (5). Different components in the substance or different hydrogen-containing groups in the same component have rich absorption spectra in the near-infrared spectral region (6). Fruits and vegetables are mainly composed of water. In addition, they also contain various vitamins, inorganic salts, and cellulose, which are rich in O-H, N-H, and C-H groups. Each molecular component has its specific absorption characteristics, which makes near-infrared spectroscopy advantageous for qualitative and quantitative detection. In terms of fruit detection, near-infrared spectroscopy has been used to evaluate the internal and external quality parameters of different types of fruits, such as apples, pears, bananas, peaches, grapes, and so forth (7–11). Similarly, near-infrared spectroscopy also shows its unique advantages in the rapid and nondestructive detection of the internal and external quality indexes of citrus. Some studies have reported the use of near infrared spectroscopy to detect early fun- gal infections in citrus. Studies have shown that near-infrared spectroscopy detection technology can be used to detect fungal infections in citrus and is an important method. Lorente and others collected the reflectance spectra of the healthy and decomposed surface parts of citrus in two spectral regions of 650 nm–1050 nm and 1000 nm–1700 nm, and adopted three different dimension reduction methods combined with linear discriminant analysis (LDA) to establish a supervised classifier (12). Finally, near-infrared spectroscopy was used for factor analysis (FA). The accuracy of partial classification of healthy and decay (diseased) citrus was 100% and 94.4%. How- ever, in this type of study, the experimental design did not distinguish the degree of decay; it only studied the characteristics of healthy and decayed samples, without considering the influence of time on the degree of decay. Ling Xia Huang and others used near-infrared spectroscopy and electronic nose technology to predict the number of days before peach decay (13). The accuracy of the best prediction model for this method was reported to be 82.26%, which is useful for the nondestructive evaluation of healthy (or edible) time before onset of peach decay. However, this prediction accuracy is not considered high enough for satisfactory decay prediction results. Clara Shenderey and others used near-infrared spectroscopy equipment to detect apple kernel mildew online, and the classification results are very accurate, indicating that using near-infrared spectroscopy to detect mildew is satisfactory (14). Yang and others studied the discriminant analysis of apple mold heart disease by using Fourier-transform near-infrared spectroscopy (FT-NIRS) and electronic nose technology combined with chemometrics (15). The discriminant accuracy of the verification set was 87.7% and 86.2% respectively, which discriminated healthy apples and apples with mold heart disease, but did not distinguish the degree of apple mold heart disease.

Near-infrared spectroscopy combined with a variety of qualitative methods is widely used in fruit detection. Choosing different modeling algorithms will affect the accuracy and stability of the model (16). Tian et al. established LDA and SVM classification models for apple nuclear mold based on full transmission spectra and established a better classification model (17). Pan et al. used near-infrared hyperspectral image detection technology to perform non-destructive classifications of apple damage time and used a linear discriminant analysis (LDA) classifier to establish apple damage time classifications, with the accuracy for this method reaching 92.86% (18). Sadjad Abasi et al. used wavelet transform denoising to preprocess spectral data and then combined it with a partial least squares (PLS) algorithm to establish a regression model and obtain a high correlation coefficient and a low root means square error (19). In this current paper, we propose to develop a qualitative discriminatory model for early fungal infections in citrus, drawing on the above modeling algorithms and approaches.

In this feasibility study, FT-NIR spectroscopy was used to identify early fungal infections in citrus fruit. The principle of identification is based on the absorbance spectra of the different fruit types measured. The change in absorbance for the different fruit quality types is mainly caused by the change in water concentration in the infected and non-infected fungal parts in the fruit samples. Due to the measurement noise of the spectra acquired by FT-NIR, the original spectra is firstly subjected to discrete wavelet transform to reduce the influence of noise, followed by a feature extraction method to reduce the dimensionality; finally, two classifiers, linear and non-linear, are used to establish different discriminant models for the spectral matrix. The evaluation of different discriminant models is mainly in terms of accuracy, sensitivity and specificity. The results of this study indicate that the use of the FT-NIR detection technique has great advantages in early fungal infections in citrus and can provide practical references for other crops in agricultural research, in addition to expanding the development of FT-NIR in related fields.

The citrus samples were purchased from a fruit market in Nanchang City, and the fungi variety included wok mandarin mold and Italian green mold. Green mold strains, which are the main culprit of citrus fungal infections, were purchased from the Beina Bio (a microbial strain research and development, production and sales group of professional institutions), and were mainly used to prepare the fungal suspension to infest the citrus. After selecting 200 orange samples with no significant difference in visual appearance, they were washed and disinfected with three different solutions (22 g/kg hypochlorite, sterile water, and 55 g/kg concentrated ethanol) for half an hour, then wiped clean with a dry rag. A hole 1 mm wide and 2.5 mm deep was drilled near the equator of each citrus sample, 100 of which were inoculated with 0.02 mL of fungal suspension at a concentration of 106 spores per mL, and the other 100 citrus samples served as a control group inoculated with 0.02 mL of sterile water, and the samples were subsequently stored in a controlled environmental chamber at 25 °C with a humidity of about 95% ± 1%. The initial phenomenon after fungal inoculation was a decay spot about 10 mm in diameter, which was hardly noticeable to the naked eye. The decay formation time of different citrus samples varied from 4 to 6 days (2,20). The early spectra of the test citrus had relatively small differences. On the 5th day of inoculation, a water-soaked area could be seen at the injection site, so the subsequent spectra were collected from the 5th day to the 13th day. Figure 1 shows citrus infected samples with a fungal suspension for five days and its control (injected with sterile water), at which time a more obvious area of water immersion existed near the borehole.

FIGURE 1: Outward images of a fungus-infected citrus (left) and a healthy citrus (right).

The Kennard-Stone (K-S) sample selection algorithm method is based on the Euclidean distance between variables, and the number of samples is selected uniformly in the feature space. For each remaining sample, the sample with the shortest distance from the selected sample is selected first, then the sample corresponding to the longest distance among these shortest distances is selected. These steps are repeated until the number of samples selected is the number of samples in the set, which is usually used for the modeling sets using spectral data (21).

This experiment intends to establish two types of qualitative discriminant models, using the K-S algorithm to divide the samples into training and test sets according to a 3:1 ratio, respectively. Table I shows the training set and test sets of the two models, as well as the citrus differentiation and sample number. Model 1 was designed to determine whether citrus was infected by fungi, and model 2 was designed to determine the degree of citrus fungal infection, which was mainly designated by 

. Among the sample types, no infection means that sterile water was injected. The 5th day was set as 

. After the 15th day, the decayed part of the citrus sample was large, and the sample had no further experimental value. A total of 100 citrus samples were selected. Due to the different growth effects of the fungus on each citrus, 103, 100, 96, and 102 citrus samples with no infection, mild infection, moderate infection, and severe infection, respectively, were selected for evaluation.

Diffuse reflection (DR) light projects onto the citrus sample, and the reflection occurs on the surface of the object in a cone-like direction—the application of this method of detection is called diffuse reflection detection. The acquisition instrument for this experiment was a Bruker multi-purpose analyzer (MPA)-type Fourier transform near-infrared spectrometer, the infrared spectra acquisition software is OPUS, and the wavenumber range is 10000 cm

, and the number of averaged scans at 32. Finally, the diffuse reflectance spectra of citrus were obtained and analyzed using the averaged spectra for each sample.

Spectral Pretreatment and Stoichiometric Analysis

Figure 2 illustrates that models were developed to discriminate healthy from decay fruits. Similarly, models were developed to also detect different levels of fruit decay. In this paper, the raw spectral data were firstly pre-processed with discrete wavelet transform (DWT) to reduce the influence of noise on modeling. Then, the data was combined with partial least squares discriminant analysis (PLS-DA) to build model 1, the use of feature extraction methods can extract useful information, reduce the spectral matrix dimensions, and improve the accuracy and stability of the identification model. Principal component analysis (PCA) is used to reduce the dimensionality of preprocessed FT-NIR spectral data, successive projection algorithm (SPA), uninformative variable elimination-successive projection algorithm (UVE-SPA), and competitive adaptive reweighted sampling (CARS) three feature extraction algorithms are used to select the feature variables of the citrus fungal infection degree discriminant model. Finally, linear discriminant analysis (LDA) and support vector machine (SVM), combined with different feature extraction methods, were applied to build a qualitative discriminant model 2. This process was implemented using MATLAB 2014b and Unscrambler X10.4 software.

FIGURE 2: Flow chart of the main steps of data analysis for two types of qualitative models.

Wavelet transform (WT) is a powerful tool based on the Fourier transform, which can transform the signal in both time and frequency domains. In spectral denoising and smoothing, the discrete wavelet transform (DWT) is generally used. In spectral analysis, the most commonly used wavelet functions are mainly Haar wavelet, Daubechies (dbN) wavelet, Coiflet wavelet and Symlets wavelet—the db4 wavelet basis was chosen after comparison (22)—this expression is shown in equation 1:

 is the translation parameter, and is the wavelet basis.

Principal component analysis (PCA) is an analysis method that achieves data reduction by projection, where the dimensionality reduction allows a few new variables to replace the many original variables in a linear combination, so that a large data space can be converted into a small dimensional factor space (23–24). The new spectral data of this small factor space are termed 

 (PC), with the largest variance of the projection on the first coordinate (PC1) and the second largest variance on the second coordinate (PC2), according to which the new variables of the transformation are orthogonal and uncorrelated with each other. The matrix after PCA can be expressed as shown in equation 2, where 

 is the score matrix composed of PCs and 

The uninformative variable elimination (UVE) algorithm is a feature variable extraction method based on PLS regression coefficients, which are used to measure the significance of the spectral bands. In the program calculation, a set of random variables of the same dimension as the spectral matrix will be randomly generated as a reference, and the stability value and threshold will be used to evaluate the validity of each variable, and variables with absolute values of stability less than the threshold will be eliminated (25). The equation 3 is shown as follows:

) denotes the mean of the regression coefficient, Std (

) is the standard deviation of the regression coefficient (

th variable in the calibration set sample, 

 is the number of input variables. The decision of whether a variable is excluded or not is based on the size of 

After extracting variables by UVE, there are still many variables left. In the UVE-SPA method, SPA is used to select fewer variables by reducing the number of variables by vector projection, and by selecting the least redundancy and covariance of selected variables (26,27).

Competitive adaptive reweighted sampling (CARS) is a feature extraction method that uses adaptive reweighting sampling techniques and exponential functions for variables with large absolute values of regression coefficients in the PLS model. The optimal subset of 

 variables with the minimum root mean square error (RMSE) of interaction is selected by the Monte Carlo method using an iterative competition approach (28,29). The main steps are as follows: 1) Monte Carlo sampling N times, randomly selecting samples as the calibration set and building the PLS model; 2) forcing the variables to be screened out using the exponential decay function; 3) further competitive selection of variables using the adaptive reweighted sampling technique; 4) calculating the RMSECV of each new variable subset model and selecting the smallest one as the optimal extracted variable.

Partial least squares discriminant analysis (PLS-DA) is a supervised qualitative classification method based on PLS regression, where the independent variable 

 is the spectral matrix and the concentration variable 

 is replaced by a known sample class designator (30). In this paper, the spectral matrix is trained and the PLS prediction of the training sample is discriminated correctly if it is within ±0.5 of the given sample category designator, a value that is real and ideally close to the value of the actual sample category designator.

LDA is a generalization of Fisher’s linear discriminant for finding linear combinations of features characterizing or separating two or more classes, and is a good analytical method in spectral processing. The LDA algorithm can be formulated as the following optimization problem:

The support vector machine (SVM) algorithm is a supervised nonlinear learning method that is most widely used in pattern recognition. SVMs usually find equidistant vertical classification surfaces between two classes and compute the maximum distance separating samples from different classes. The kernel function and the kernel parameters are the two main computed parameters when it comes to SVM classification, and determining the penalty factor (c) and the kernel parameter (g) is the key to model building (32). The radial basis kernel function (RBF) usually has better prediction results, a more mature theoretical system, is more widely used, and is considered to be the first choice for SVM modeling. Radial basic function (RBF) is described as follows:

Data Analysis Flow Chart and Model Evaluation

For NIR spectral data, more sample data are available for the model-building to distinguish between infected and healthy citrus and different levels of infection. Some spectral data wavenumber ranges do not provide useful predictive information. Extracting useful information can greatly improve the accuracy of the model. In this study, two recognition models were designed to determine whether the citrus has a fungal infection and the degree of infection. The raw spectral signals of different types of citrus were obtained by FT-NIR, and the raw signals were subsequently pre-processed using discrete wavelet transform. The features of the two qualitative discriminant models were extracted separately using PCA, and three variable selection algorithms were used to select the feature variables of the second discriminant model. Finally, LDA and SVM, combined with different feature extraction and variable selection methods were used to establish two qualitative discriminant models. The model evaluation parameters are the correct rate of identification, the correlation coefficient, and the root means square error of the modeling and prediction sets of the two types of qualitative models. The higher the correct rate, the larger the correlation coefficient and the smaller the root mean square error, and thus, the better the modeling effect (21). A flow chart of the main steps of the data analysis of the two types of qualitative models is shown in Figure 2.

Fourier Transform Near-Infrared Spectra of Citrus With or Without Disease and Citrus with Different Disease Degrees

Near-Infrared Mean Spectra of Healthy and Decay Citrus Samples 

Figure 3 shows the original spectra of decay citrus and healthy citrus. The two spectral trends are similar. Overall, the reflection absorbance of healthy citrus is higher than that of decay citrus, and spectra exhibit three main peaks. The range of 5200 cm

 is related to the O-H stretching, O-H band binding and deformation of the representative H-O-H of starch content. A peak can be observed at 6988 cm

 of O-H stretch 1st overtone for the C=O stretch 3rd overtone. The peak in the range of 8500 cm

 is related to the second overtone of the aromatic structure (33,34). Overall, the absorbance of citrus decreases as the wavenumber decreases. The absorbance based on diffuse reflection spectra can be used for modeling and analysis of healthy and decay citrus samples.

FIGURE 3: Near-infrared mean absorbance spectra of sound (healthy) and decay (diseased) citrus.

Average Near-Infrared Spectra of Citrus Samples with Different Degrees of Decay

Figure 4 shows the average NIR spectra of citrus samples with different degrees of decay, from which it can be easily seen that overall healthy citrus has the highest absorbance, followed by light decay, moderate decay, and severe decay. As time went by, the chemical properties of citrus in the infected area changed greatly. Based on the differences in spectral information, a series of modeling methods will be used to qualitatively determine the degree of fungal decay.

FIGURE 4: Near-infrared mean spectra of citrus samples with different degrees of decay.

Establishment of Qualitative Discrimination Model for Healthy and Decay Citrus Samples

Due to the existence of noise in the spectral information, the DWT was used to process the original spectra, and then the PLS-DA model was established separately with scale size lev = 3, threshold selection method of Heursure function, wavelet basis function of db4, signal-to-noise ratio S/N = 24.0949, and RMS = 0.04289. Table II shows the PLS-DA model based on the full spectral information without preprocessing and DWT processing, which shows that the effect of the spectral matrix after discrete wavelet transform is better than the original spectral information, that the correctness of both the training and prediction sets is 100%, and that the deviation is only 0.0047 with low root mean square error. This indicates that the discrete wavelet transform can better reduce the effect of noise on data processing, so the discrete wavelet transform will be used first in the subsequent modeling of qualitative discrimination of fungal infection degree to optimize the model.

Selection of Characteristic Variables of Fourier Near-Infrared Spectroscopy

The Fourier NIR contains 3111 wavenumbers with more information. To reduce the experimental error and the influence from noise, the selection of suitable feature variables is crucial for later modeling. Too much selection of feature variables and a large amount of irrelevant information will affect the accuracy and stability of the model; on the other hand, insufficient selection of feature variables may lead to the loss of useful information, thus failing to build an accurate model. Therefore, in this paper, discrete wavelet transform is first used to preprocess the spectra, and then PCA, SPA, UVE-SPA, and CARS are used to select characteristic variables for the original data. Reduce data dimension, remove useless information, extract effective spectral variable information.

PCA is one of the commonly used methods for dimensionality reduction of data. The transformed variable space is a linear combination of the variables in the original variable space and represents the vast majority of information (24). Figure 5 shows the wavelength variable screening process diagram of the PCA algorithm. The maximum number of principal component factors is set to 20. When eight principal components are selected, the new variable space can represent more than 99.8% of the information of the original spectra.

FIGURE 5: The process of spectra variables selection for PCA (PC vs. variance contribution %).

Feature Wavelength Screening Based on SPA

The successive projections algorithm (SPA) is a forward cyclic selection method. The optimal sample set is obtained by calculating the root mean square error of multiple linear regression models of different sample subsets, and the effective wavelength of the minimum redundancy and collinearity is selected by vector projection (26). This model is established in MATLAB2014b, and the minimum and maximum selection variables are set to be 20 and 150. Figure 6 shows the results after the wavelength screening of the sample spectra based on the SPA algorithm, where the horizontal and vertical coordinates represent the wavelength and the average spectral reflectance, respectively. 82 wavenumbers were selected in the original spectra using the SPA algorithm, and only 2.64% of the data were used compared with the original spectra, which can eliminate redundant information to a certain extent and simplify the model, and the variables selected by SPA will be used subsequently for model building.

FIGURE 6: SPA selected wavenumber results (orange circles).

Feature Wavelength Screening Based on UVE-SPA

The uninformed variable elimination (UVE) is a band screening method based on the partial least squares (PLS) regression coefficient, and the regression coefficient is used to measure the significance of the band (27). When the UVE algorithm is used alone, although some characteristic wavelength points that have no information are eliminated, there will be a problem of collinearity among the selected characteristic wavelength points. Similarly, when the SPA algorithm is used to extract the characteristic wavelengths, although the collinearity between the characteristic wavelength points is reduced, the selected wavelength subset is likely to contain some uninformative and even interfering wavelengths (35). Therefore, the UVE algorithm and SPA algorithm are combined in this experiment to realize the complementary advantages of the two. UVE generally preserves more wavenumbers. Firstly, 1837 wavenumbers are selected by UVE, and then 97 feature variables are selected by SPA to obtain the most useful characteristic variables. If only using SPA modeling, the UVE-SPA model established will be more conducive to the establishment of a robust discrimination model (Figure 7).

FIGURE 7: UVE-SPA wavelength variable selection results. (Red star marks the point of final number of selected variables included in the model.)

Feature Wavelength Screening Based on CARS

Competitive adaptive reweighted sampling (CARS) is a feature variable selection algorithm combining the Monte Carlo sampling method with partial least squares (PLS) regression coefficient (28). Adaptive reweighted sampling is adopted for CARS. A portion of samples is mainly extracted from the training set for PLS modeling analysis. Through repeated calculation, the subset with the smallest interactive root means square error (RMSECV) is selected as the selected wavelength variable (29). Figure 8 shows the relationship between the number of wavelength variables, interactive root means square error, the regression coefficient of wavelength variables, and the change of sampling times in the CARS algorithm. As can be seen from Figure 8a, the number of retained wavelengths decreases with the increase of sampling times, and then gradually tends to zero. This phenomenon shows the process from rough selection to selection of wavelength variables by CARS algorithm most vividly. Figure 8b shows the trend diagram of the relationship between RMSECV and the sampling times. When the sampling times gradually increase from 1 to 43, the value of RMSECV drops gradually, and the minimum value is obtained at the sampling times of 43. When the sampling times are greater than 43, the RMSECV value increases gradually. The results showed that the CARS algorithm could screen out the information that interfered with the spectra of citrus fungal infection degree when the sampling times were less than 43, while the important information related to citrus fungal infection degree was screened out when the sampling times were greater than 43. Figure 8c shows the relationship between the regression coefficient of the wavelength variable and the number of sampling times. The position of “*” represents the sampling times corresponding to the minimum RMSECV value, and the minimum RMSECV value is 43. Through these variables, most of the information of the whole sample can be replaced and the high precision model can be established.

FIGURE 8: CARS variable screening stability result graph. (a) Select the number of wavelength variables, (b) the interactive root mean square error, and (c) the regression coefficient of wavelength variables.

Different Characteristic Variable Selection Methods Combined with Model Establishment of LDA and SVM 

The discriminant models using linear discriminant analysis (LDA) and support vector machine (SVM) classifiers combined with four feature variable selection methods, respectively, are given in Table III, with the parameters, the number of selected variables and the correctness rate. The comparison shows that the SVM model has better performance than the LDA model. The CARS-SVM model based on the RBF kernel function works best with 82 number of selected feature variables, an average correct rate of 100% in the training set and 97% in the prediction set, and the PCA-LDA model works the worst with an average correct rate of 75.33% in the prediction set. This may be due to the ability of support vector machines to identify nonlinear and high-dimensional data and to find the best discriminative boundary by mapping from low to high dimensional space, thus improving the prediction ability and increasing the correct classification rate (29). Among these feature variable selection methods, CARS works the best for both LDA and SVM, and by reweighing the sampling and exponential decay function, CARS can effectively eliminate variables that contribute little to the model, while other variable selection methods more or less ignore some irrelevant information or remove useful information. In summary, this feature extraction method can improve the operation speed and correctness of the discriminative model on FT-NIR spectral analysis.

In this paper, citrus is the research object, and the feasibility of FT-NIR for the detection of early fungal infections in citrus is studied. Firstly, to discriminate whether citrus is fungal infected or not, the PLS-DA model was built based on the original data and DWT, and FT-NIR after DWT treatment showed good potential with 100% correctness in both training and prediction sets. Based on this, the subsequent modeling was also first processed with DWT and then combined with PCA, SPA, UVE-SPA, and CARS, to select feature variables to achieve dimensionality reduction of the spectral matrix. Finally, two qualitative models, LDA and SVM, were used to establish the identification model of the degree of citrus fungal infection. The results showed that the SVM discriminative model based on radial basis function was better than LDA. In addition, the results also showed that the accuracy and robustness of the discriminative model could be well improved by using the feature variable selection method. DWT-CARS- SVM was the best, with 100% and 97% correctness in the training and prediction sets, respectively. Therefore, FT-NIR can be a practical tool for rapid, nondestructive, and accurate detection of early fungal infections in citrus. However, the time interval for identification of infection level in this paper was 3 days, and developing more accurate detection algorithms as well as detection accuracy remains a big challenge, so further research is needed to improve the model.

This work was supported by the National Natural Science Foundation of China (Grant number: 31760344); National Science and Technology Award Reserve Cultivation Project of Jiangxi Province (Grant number: 20192AEI91007).

Maopeng Li: Writing-Original Draft, Data Curation, Methodology. Yande Liu: Investigation, Validation, Formal Analysis. Jun Hu: Writing (Review and Editing), Validation, Software. Chengtao Su: Resources. Zhen Xu: Resources. Huizhen Cui: Resources

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

(1) Itakura, K.; Saito, Y.; Suzuki, T.; Kondo, N.; Hosoi, F. Estimation of Citrus Maturity with Fluorescence Spectroscopy Using Deep Learning. 

 (1), 2. DOI: 10.3390/horticulturae5010002

(2) Ghanei Ghooshkhaneh, N.; Golzarian, M. R.; Mamarabadi, M. Detection and Classification of Citrus Green Mold Caused by Penicillium digitatum Using Multispectral Imaging. 

 (9), 3542–3550. DOI: 10.1002/jsfa.8865

(3) Chu, B. Q.; Zhang, H. L.; Luo, W.; He, Y. 

Spectrosc. Spectral Anal. (Beijing, China)

(4) Tian, X.; Fan, S. X.; Huang, W. Q.; Wang, Z. L; Li, J. B. Detection of Early Decay on Citrus Using Hyperspectral Transmittance Imaging Technology Coupled with Principal Component Analysis and Improved Watershed Segmentation Algorithms. 

, 111071. DOI: 10.1016/j.postharvbio.2019.111071

(5) Simeone, M. L. F.; Parrella, R. A. C.; Schaffert, R. E.; Damasceno, C. M. B.; Leal, M. C. B.; Pasqiuni, C. Near Infrared Spectroscopy Determination of Sucrose, Glucose and Fructose in Sweet Sorghum Juice. 

, 125–130. DOI: 10.1016/j.microc.2017.05.020

(6) Zhang, C.; Su, J. H. Application of Near Infrared Spectroscopy to the Analysis and Fast Quality Assessment of Traditional Chinese Medicinal Products. 

 (3), 182–192. DOI: 10.1016/j.apsb.2014.04.001

(7) Pissard, A.; Marques, E. J. N.; Dardenne, P.; Lateur, M.; Pasqiuni, C.; Pimentel, M. F.; Pierna, J. A. F.; Baeten, V. Evaluation of a Handheld Ultra-compact NIR Spectrometer for Rapid and Non-destructive Determination of Apple Fruit Quality. 

, 111375. DOI: 10.1016/j.postharvbio.2020.111375

(8) Li, J. B.; Wang, Q. Y.; Xu, L.; Tian, X.; Xia, Y. Fan, S. X. Comparison and Optimization of Models for Determination of Sugar Content in Pear by Portable Vis-NIR Spectroscopy Coupled with Wavelength Selection Al- gorithm. 

 (1), 12–22. DOI: 10.1007/s12161-018-1326-7

(9) Saputro, A. H.; Juansyah, S. D.; Handayani, W. Banana (Musa sp.) Maturity Prediction System Based on Chlorophyll Content Using Visible-NIR Imaging. In 

2018 International Conference on Signals and Systems (ICSig- Sys

), Bali, Indonesia, May 1–3, 2018; IEEE; pp 64–68. DOI: 10.1109/ICSIGSYS.2018.8373569

(10) Li, L.; Wu, Y.; Li, L.; Huang, B. Q. Rapid Detecting SSC and TAC of Peaches Based on NIR Spectroscopy. In 

2017 2nd IEEE International Conference on Computational Intelligence and Applications (ICCIA)

, Beijing, Sept 8–11, 2017; IEEE; pp 312–317. DOI: 10.1109/CIAPP.2017.8167229

(11) Daniels, A. J.; Poblete-Echeverría, C.; Opara, U. L.; Nieuwoudt, H. Measuring Internal Maturity Parameters Contactless on Intact Table Grape Bunches Using NIR Spectroscopy. 

(12) Lorente, D.; Escandell-Montero, P.; Cubero, S.; Gómez-Sanchiset, J.; Blasco, J. Visible–NIR Reflectance Spectroscopy and Manifold Learning Methods Applied to the Detection of Fungal Infections on Citrus Fruit. 

, 17–24. DOI: 10.1016/j.jfoodeng.2015.04.010

(13) Huang, L.; Meng, L.; Zhu, N.; Wu, D. A Primary Study on Forecasting the Days before Decay of Peach Fruit Using Near-Infrared Spectroscopy and Electronic Nose Techniques. 

, 104–112. DOI: 10.1016/j.postharvbio.2017.07.014

(14) Shenderey, C.; Shmulevich, I.; Alchanatis, V.; Haim, E.; Aharon, H.; Viacheslav, O.; Susan, L.; Ruth, B. A.; Ze’ev, S. NIRS Detection of Moldy Core in Apples. 

 (1), 79–86. DOI: 10.1007/s11947-009-0256-1

(15) Yang, C. Y.; Yuan, H. F.; Ma, H. L.; Ren, Y. M.; Ren, X. L. Nondestructive Detection of Apple Moldy Core Based on FT-NIR and Electronic Nose Technology. 

 (7), 211–216. DOI: 10.13995/j.cnki.11-1802/ts.025671

(16) Tian, X.; Fan, S. X.; Li, J. B.; Xia, Y.; Huang, W. Q.; Zhao, C. J. Comparison and Optimization of Models for SSC On-line Determination of Intact Apple Using Efficient Spectrum Optimization and Variable Selection Algorithm. 

, 102979. DOI: 10.1016/j.infrared.2019.102979

(17) Tian, X.; Wang, Q. Y.; Huang, W. Q.; Fan, S. X.; Li, J. B. Online Detection of Apples with Moldy Core Using the Vis/NIR Full-Transmittance Spectra. 

, 111269. DOI: 10.1016/j.postharvbio.2020.111269

(18) Pan, X. Y.; Sun, L. J.; Li, Y. S.; Che, W. K.; Ji, Y. M.; Li, J. L.; Li, J.; Xie, X.; Xu, Y. T. Non-destructive Classification of Apple Bruising Time Based on Visible and Near-Infrared Hyperspectral Imaging. 

 (4), 1709–1718. DOI: 10.1002/jsfa.9360

(19) Abasi, S.; Minaei, S.; Jamshidi, B.; Fathi, D.; Khoshtaghaza, M. H. Rapid Measurement of Apple Quality Parameters Using Wavelet Denoising Transform with Vis/NIR Analysis. 

, 7–13. DOI: 10.1016/j.scienta.2019.02.085

(20) Gómez-Sanchis, J.; Martín-Guerrero, J. D.; Soria-Olivas, E.; Martínez-Sober, M.; Magdalena-Benedito, R.; Blasco, J. Detecting Rottenness Caused by Penicillium Genus Fungi in Citrus Fruits Using Machine Learning Techniques. 

 (1), 780–785. DOI: 10.1016/j.eswa.2011.07.073

(21) Kennard, R. W.; Stone, L. A. Computer Aided Design of Experiments. 

 (1), 137–148. DOI: 10.1080/00401706.1969.10490666

(22) Zhang, J.; Yuan, S.; Zhang, J. Detection of Sugar Degree in Strawberry Based on Wavelet Transform-Genetic Algorithm-Partial Least Squares. 

 (1), 111–116. DOI: 10.13526/j.issn.1006-6144.2020.01.020

(23) Liu, Y. D.; Li, P. M.; Hu, J.; Xu, Z.; Cui, H. Z. Identification of Coffee-Bean Varieties Using Terahertz Detection Technology. 

 (16), 525–530. DOI: 10.3788/LOP202158.1630002

(24) Qin, B. Y.; Li, Z.; Luo, Z.; Li, Y.; Zhang, H. Terahertz Time-Domain Spectroscopy Combined with PCA-CFSFDP Applied for Pesticide Detection. 

 (7), 1–12. DOI: 10.1007/s11082-017-1080-x

Articles

We examine the feasibility of FT-NIR for the detection of early fungal infections in citrus.

Detection of the Early Fungal  Infection of Citrus by Fourier Transform Near-Infrared Spectra

Honey peaches can bruise during harvesting, handling, storage, transportation, and distribution. In this study, the spectral range used was 400–1100 nm, and we extracted the RGB and HSI color space characteristics of the images. After principal component analysis (PCA) of the original data, the gray histogram features of the PC1 images were extracted. Partial least squares qualitative discriminant analysis (PLS-DA) and extreme learning machine (ELM) discriminant models were established. Among the 38 color features, the PLS-DA and ELM models had a high rate of misclassification, and the best classification accuracy was 74.29%. When extracting the spectral information of the bruised sample to build the model, the highest classification accuracy was 92.86% for the 176 characteristic wavelength points of the full band. In contrast, only 40 wavelength bands were used after selecting the genetic algorithm’s valid information. The classification accuracy of the PLS-DA model was 100%, which is because the softening and browning of the peach was not apparent after early bruising. However, the changes in the tissue’s thermal properties caused by internal defects are expressed in the internal spectrum. Therefore, the shortwave NIR hyperspectral imaging technique’s spectral information can detect the early bruising of peaches.

Honey peaches are one of the most critical economic fruits in the world. China has the largest peach plant area and the largest production scale compared to other countries such as Italy, Greece, Spain, and the United States (1). From a nutritional perspective, honey peaches are a good source of dietary fiber, vitamin B, vitamin C, carbohydrates, minerals, organic acids, folic acid, and dietary antioxidants. The protein content of honey peaches is twice that of apples and grapes and seven times higher than pears. The iron content is three times higher than apples and five times higher than pears (2). The presence of bruise defects is one of the most influential factors in the quality and price of a honey peach. In general, customers pick up a honey peach and judge its quality by external attributes such as shape, color, and size. A moderate amount of bruising often deters consumers from purchasing a peach, which can lead to postharvest loss and an overall decrease in profits across the industry. Moreover, honey peaches affected by bruises will tend to ferment, decay, and infect other non-bruised peaches after damage. However, naked eye observations are limited in evaluating peach bruising. Moreover, efficiency and accuracy would be significantly reduced after continuous manual inspection. Therefore, it is essential to develop a rapid, non-contact detection technique to identify the honey peach bruising and determine its bruise degree. By sorting honey peaches following their bruise degrees, it would result in higher quality peaches hitting the market, which means better prices and less food waste, increasing profits (3,4).

Hyperspectral imaging is an emerging technique in recent years that can provide both spatial and spectral information. During hyperspectral imaging analysis, the spectra are used as a fingerprint to identify and measure the vibration in energy absorption that depends on the sample’s chemical composition and moisture content. In general, machine vision technology based on a red, green, and blue (RGB) color camera is more effective for detecting those bruises (5). Hyperspectral imaging also has potential in detecting bruises in various fruits, including peaches. Hyperspectral imaging integrates conventional imaging and spectroscopy to provide both the spatial and spectral information of an object. Each hyperspectral image contains a large amount of information that can be analyzed to characterize the object more reliably than the traditional machine vision technology (6–10). A few characteristic wavelength images could be selected to develop rapid multispectral imaging systems for fast bruise detection. It is advanced and efficient to detect non-obvious bruises of samples and assess internal attributes of fruits. Keresztes and others used hyperspectral imaging to detect the varieties and bruises of apples, and the correct rate of the classification of varieties and bruises reached 96% and 90.1%, respectively (11). Lopez-Maestresalas and others studied potato bruises in the visible-near-infrared (vis-NIR) and shortwave NIR wavelength ranges, with a classification accuracy rate of 98.56%, showing that early bruises can be detected on food items (12). Vetrekar and others verified the feasibility of nondestructive testing of apple, heart fruit, and guava fruit bruises in the wavelength range of 400–1000 nm (13). Zhang Baohua and others used hyperspectral imaging combined with the minimum noise separation algorithm (MNF) to detect minor bruises with an accuracy rate of 97.1% (14). Wu Longguo and others used hyperspectral imaging to study the various defects of Lingwu Changzao jujubes and insect eyes, and the accuracy rate of various defects reached more than 86% (15). Jiangbo Li and others used longwave and shortwave NIR hyperspectroscopy combined with an improved watershed algorithm to analyze early peach bruising and obtained excellent results, laying a foundation for the study of early peach bruising (16). These reports show that NIR hyperspectral imaging combined with the appropriate detection algorithms can effectively identify bruised fruit.

Partial least squares (PLS) regression is the most widely used algorithm in the process of solving the linear relationship between hyperspectral imaging and target data (17). Extreme learning machine (ELM) has many unique advantages in solving small sample, nonlinear, and high-dimensional pattern recognition. ELM has a reduced computational complexity (18). Variable selection is another critical factor affecting the accuracy of a prediction model. To improve the accuracy and robustness of the prediction model, the compensation model of biological or environment variability was built with variable selection tactics. The selection of characteristic wavelength can eliminate the collinearity, redundancy, and noise in the full wavelength spectrum, thus improving the prediction accuracy, which involves many variable selection algorithms. However, the results obtained from different variable selection algorithms are usually different because of their varied principles and applications (19,20).

The accuracy of fruit bruise detection is influenced by many factors, such as the variety of bruises, location, severity, fruit variety, and storage time. Some can be detected directly by image segmentation, but some of the minor bruises cannot be observed by machine vision technology. It is necessary to analyze the mechanism of fruit bruising, taking into account changes in the internal spectrum of the fruit in question. In previous studies, it has been suggested that shortwave NIR is sufficient to distinguish the early bruising of peaches (21). This study started with using simplified equipment. It only used the hyperspectral shortwave camera to collect sample information and further analyze the results (22,23). The aims of this study were to: 1) extract the characteristics of the image of the sample, including the RGB, HSI color characteristics, and gray histogram statistical characteristics of the model; 2) compare the spectral features of sound peaches and bruised peaches, and mechanically find out the differences between sound peaches and bruised peaches; 3) compare the six detection results produced by the two algorithms using image feature information, full-band spectral information, and the influential band of the spectra selected by the genetic algorithm (GA) as input values of the prediction model, respectively; and 4) draw conclusions for distinguishing bruised peaches from uninjured tissues and provide a theoretical basis for online fruit bruise detection.

The experimental sample selected in this study was the honey peach, and the peach variety was Yangshan. Figure 1 shows a picture of the experimental sample. To prevent other factors from affecting the experimental results, the samples’ size and weight remained identical. The equatorial diameter of the honey peach was 80 ± 5 mm; all models were carefully selected before the experiment to ensure that the samples’ appearance was free from defects and damages. The bruised parts were chosen near the equator of the honey peach. The number of experimental models was 210, of which 120 were injured and 90 were undamaged. A steel ball was used to obtain the bump sample and form a bruise. The mass of the steel ball was 33 g, the diameter of the steel ball was 20 mm, and it was placed 160 mm above the surface of the honey peach. Then, the free-fall motion was performed, and a bump sample was formed by hitting the equator of the honey peach. Finally, the 210 pieces were placed in an environment with a room temperature of 24 °C and relative humidity of 65%, and the image was collected 1 h later. The purpose of standing for 1 h was to keep the sample temperature consistent with the ambient temperature. Figure 2 shows a schematic diagram of the simulated impact (16).

FIGURE 1: Picture of the experimental sample.

FIGURE 2: Schematic diagram of the bruise simulation.

The left of Figure 1 shows the intact sample, and the right shows the slight bump in the piece. The total number of samples is 140, 70 samples are used for modeling and the remaining 70 samples are used to validate the model. The number of pieces in the prediction set was used to evaluate the model.

Hyperspectral Imaging System and Data Acquisition

The hyperspectral imaging system included a light source, an imaging spectrometer, a charge-coupled device (CCD) camera, a lens, a displacement platform, a controller, a dark box, and a computer. The spectral resolution was 2.8 nm, and the spectral range was from 380 to 1080 nm. As shown in Figure 3, a schematic diagram of the structure of the hyperspectral imaging system is shown. The camera exposure time was 20 ms, the interval of spectral radiation energy was approximately 2.4 nm; the platform travel speed was approximately 15 mm/s. Because the image acquisition process was susceptible to the effects of ambient light, the entire imaging system was placed in a confined space to eliminate the influence of ambient light, which prevented the effects of natural light and improved the quality of the acquired data.

FIGURE 3: Diagram of the hyperspectral imaging system.

Because of the spatial intensity change of the halogen lamp on the acquisition plane, a dark current also existed in the CCD camera. These factors interfered with specific wavebands and affected the quality of the image. Therefore, before data processing, the data needed to be corrected in black and white. The calibration process was completed on the SpecView software. First, the lens was aimed at the whiteboard before the whiteboard button was clicked to collect the whiteboard image and background information. To prevent the lens from rotating and causing the focal length to change, the lens was held sideways to screw on the lens cover. Once that step was completed, the blackboard button was clicked to collect the blackboard image. The black and white correction was based on equation [1] below:

 is the all black calibrated image data; 

 is the all white calibrated image data; and 

Principal component analysis (PCA) enhances the information content of hyperspectral images and has the advantages of isolating noisy signals and reducing the dimensionality of the data. It is possible to comprise previously redundant information into manually irrelevant data, while at the same time replacing all of the spectral information with less information in a practical sense. The algorithm works on the following principle: suppose a given training data set contains 

 and the covariance matrix of the training set is as follows:

The eigenvalue of Σ represents the distribution variance of the sample on the eigenvector.

 eigenvectors of Σ and sort them according to the eigenvalues. The dimensionality reduction feature subspace is expressed as 

The GA has its origins in the study of computer simulations of biological systems. It is a stochastic global search and optimization method developed to mimic natural organisms' evolutionary mechanisms, drawing on Darwin's theory and Mendel's theory of genetics. It is essentially an efficient, parallel, and global search method that automatically acquires and accumulates knowledge about the search space during the search process and adaptively controls the search process to find the best solution. This study uses the genetic algorithm to select spectrally valid information, reducing the number of model calculations while improving model detection accuracy.

Partial Least Squares Discriminant Analysis (PLS-DA)

The partial least squares linear discrimination analysis (PLS-DA) method is a regression model that can establish the relationship between sample categorical variables and spectral characteristics. Sample classification variables are used instead of sample variables. The linear relationship between the statistical classification variables and the spectral matrix was used to establish a qualitative discrimination model. The calculation steps of peach spectrum 

, respectively. The linear regression equations of 

 is the correlation matrix. The score array 

 of the unknown sample is obtained from the spectral array 

. Then, the collision time of unknown samples can be calculated according to equation [9]:

The flow chart of calculation during modeling is shown in Figure 4, and the corresponding matrix factors 

FIGURE 4: The modeling calculation flowchart.

During the prediction period, the predicted value of unknown samples can be calculated according to equation [10]:

ELM is a single hidden layer feed forward network with random input weight selection and output weight estimation analysis, assuming that in the flow chart of ELM algorithm, 

 is the weight variable that connects the implicit node to the input node; 

is a weight variable that connects the implicit node to the output node; and 

 sample spectra and actual values form a matrix of 

 columns, whereas the excitation function and hidden node L are mathematically obtained by equation [11]. The predicted results can be calculated by equation [12] (29–32).

The implied node number L is determined randomly, while the weight of the β matrix is estimated by 

 to obtain a stronger generalization performance. The conventional parameter was 

 > 0, so the global optimal solution was obtained by analyzing the ELM output weight, avoiding the complicated convergence problem in the early stage (33,34).

The sample spectral feature extraction was performed with the envi4.5 software to ensure that the selected sample spectrum was representative, prevented the single-pixel point spectrum from the overall spectral difference of the sample, and failed to express the correct information of the model. The sample spectral feature was considered the average range of the area of interest. After reading the sample image, the region of interest (ROI) in the rectangular area was selected. The image data obtained by hyperspectroscopy is a “three-dimensional (3D) data block,” in which 

 represent two-dimensional plane pixel information coordinates. The third dimension is the wavelength information coordinate axis. The number of pixels in the region of interest was selected to be 100, and the average spectrum of 100 pixels was calculated. Figure 5 shows the ROI schematic diagram of the area chosen and the ROI selection process.

FIGURE 5: The region of interest selection diagram: (a) ROI schematic diagram of the selected area and (b) the ROI selection process.

To ensure that the spectra of the pieces were chosen to be representative and to prevent the scopes of individual pixels from differing significantly from the overall range of the sample, the spectral features of the pieces chosen were the average spectrum of the region of interest. After reading the sample image, a rectangular area of interest was selected. The number of pixels in the selected region of interest was 100, and the average spectrum of the 100 pixels was calculated.

Figure 6 shows a comparison of the reflection spectra of the bruised and the sound (unbruised) sample. The abscissa is the wavelength range of the spectrum, and the ordinate is the reflectivity. The image of the bruised piece at the wavelength of 810 nm was extracted, and the pseudo-color image was processed. Figure 6 contains the single-wavelength point’s grayscale image and the pseudo-color image of the bruised sample. One of the gray values increased gradually from blue to red because of the influence of the wild peach surface radian. The edge of gray value was lower than that near the middle of the room, which means that it is difficult to segment the damaged area, so only using single wavelength point of an image is difficult in identifying mild bruising (35,36).

FIGURE 6: Spectral comparison of damage and intact samples.

The average spectra of the bruised and normal regions of yellow peaches were extracted, and comparative analysis revealed that the waveforms of the two spectra had similarities, with absorption peaks at 714 nm, 815 nm, and 970 nm. This may be caused by the second overtone of the C-H functional group (37). The absorption peaks appearing at 815 nm and 970 nm wavelength points may be caused by the O-H and N-H functional groups. The spectral differences between the bruised and normal regions of yellow peaches are also related to differences in moisture content and other organic molecules (38–40).

A scatter plot of the classification of bruised and sound peach data is shown in Figure 7. Black dots indicated a good peach, and the red dots indicated the bruised peach. The results show that there was clustering between the sound and bruised samples. It can be observed in the Figure 7 that the clustering phenomenon occurs between good peaches and bruised pieces. This phenomenon indicated that differences in the internal physiological parameters between post-bruise and sound samples could be detected spectroscopically or graphically. However, there is a partial crossover between post-bruise data and sound peach data, which makes it somewhat difficult to distinguish between grades and, as a result, requires further data analysis.

FIGURE 7: Scatter plot of bruised and Soung honey peach classification.

Using the shortwave NIR wavelength range (400–1100 nm), the peach bruises on the early samples were obtained by PCA. The high spectral reflection image of the first five images, as shown in Figure 8, represent the real PC images, to make the exact comparison of the different PC image corresponding to the pseudocolor images. As a general note, the original without the hyperspectral image change contains more information than the PC images, so the PC images have a certain degree of distortion and cannot keep the characteristics of the original image. As shown in Figure 8, the PC2 and PC4 images reveal no useful information. However, PC3 and PC5 show the bruised image contrast between the intact and damage tissue better; on the downside though, PC3 and PC5 contain segmentation errors because of the excessive area. PC3 also has a “false positive” error. Although PC1 is fuzzy, it retains the information that the image needs to be expressed. Therefore, the PC1 image was selected for the feature extraction.

FIGURE 8: Pictures of principal component analysis.

The main image features commonly used are color features, texture features, shape features, and spatial relationship features. The image features in this study refer to the color features of an image. After the fruit is bruised, the bruised area gradually changes color over time, becoming darker, and finally, black. Therefore, of all the image features, the color features are better suited to characterize the changes in the surface of the fruit after a bruise. Because of the change in color after a bruise, this change can be represented in the grayscale histogram of the image. In this study, color and gray histogram statistical features were chosen to characterize the sample image after bruising. The grayscale histogram statistical features refer to the occurrence times of corresponding pixel points in each cell of the grayscale image. The grayscale value range of the grayscale image was 0–255, and the interval selected in this study was eight. The interval range of 0–255 was divided into 32 cells, and the occurrence times of the corresponding pixel points in each cell were counted. It is important to note that in the gray-scale histogram of a multispectral, not all PCs have the same contrast (for example, in PC3, the sound and bruised areas have similar gray levels, so if the damaged areas with higher gray values are successfully segmented based on the PC3 images, this will result in a large number of sound tissues with lower gray values being misclassified).

As shown in Figure 9, the gray histogram of the PC1 image was extracted; the gray histogram was also an essential feature of the image, reflecting the relationship between the frequency of each gray level pixel and the gray level of an image. The abscissa was the interval number, which was divided into 32. As the interval number increased, the color components became lighter, and the smaller the interval value was, the darker the color was. The ordinate is the number of pixels. Figure 9 shows the bruising, sound, and transfer area. The larger the bruising size is and the higher the pixel point is, the more sensitive the bruising is in image statistics.

FIGURE 9: Gray histogram statistical feature extraction process.

Early Bruise Analysis of Peach with Hyperspectral Image Information

In this study, two color features, RGB and HSI, and 32 features from the sample’s grayscale histogram statistics, were extracted. The RGB color space is comprised of red, green, and blue components. In the RGB color space, the color of arbitrary 

R, G, and B are the three color components of red, green, and blue, respectively, and the percentages of the three colors. Other colors in the RGB color space can be expressed through equation [15]. Extract the three-channel component of the whole sample, and then calculate the average value of the three-channel element respectively, as shown in the following formula:

 are the mean values of the colors of the red, green, and blue channels, respectively; 

 is the number of pixels; and red, green, and blue are the color values of the three channels, respectively. Equation [16] can calculate the mean value of the color component of the RGB three-channel of the image and use it as the final modeling data. However, 

 can realize the conversion from RGB color space through equation [17].

A total of 38 valid features were input, and the average values of the three channel components of each color space, 

, were selected for modelling. All color features needed to be synchronized prior to building the detection model to reduce differences between the inputs. With 38 color features as the inputs, the bruised peaches set to 1, and the good peaches set to 0, there were 70 ones and 140 zeros in the total data. A regression model with qualitative discrimination between a matrix of 210 multiplied by 38 and 210 set values, which in turn contained the PLS and ELM models. Of the 210 samples, 140 of them were randomly selected as the modeling set. The remaining 70 pieces were used as the prediction set. Table I shows the early bruise area discrimination results for honey peaches based on the hyperspectral water image information, with a lower false positive rate that indicates a better model. The PLS discriminant model results show that 18 prediction sets were misclassified with a classification accuracy of 74.29%. The ELM discriminant model for the hardlim function had 26 misjudgments and only 62.86% were correct, the sigmoidal function has 22 misjudgments, the correct rate of classification is 68.57%, and the sine function was only 72.86% accurate for classification.

Figure 10 shows the effect diagram of image feature modeling: the red represents the modeling set, and the blue represents the prediction set. The threshold 

 was set at 0.5. When the value was less than 0.5, it was judged as a sound honey peach; when the value was more significant than 0.5, it was regarded as a bruised honey peach. The PLS-DA model established has a better classification effect than the ELM model. The RMSEC root mean square error of the modeling set (which was PLS-DA) was 0.4711, the correlation coefficient of the modeling set was 0.3122, the RMSEP of the prediction set was 0.4718, and the correlation coefficient of the prediction set was 0.2880.

FIGURE 10: Plot of PLS-DA image feature modelling results.

Spectral features in the 400–1100 nm wavelength range contain 176 characteristic wavelength points to better predict honey peach bruising early on. The genetic algorithm (GA) has a unique advantage in choosing its useful spectral bands, as shown in Figure 11. GA selected the adequate information of spectra, with a total of 40 useful wavelength points: 120, 81, 24, 30, 83, 175, 32, 80, 82, 84, 23, 119, 121, 85, 102, 31, 79, 25, 11, 101, 122, 22, 174, 12, 76, 118, 74, 77, 18, 21, 28, 143, 29, 124, 33, 78, 86, 103, and 104.

FIGURE 11: Genetic algorithm (GA) filtering spectrally valid information.

As shown in Table II, the early discriminant model for honey peach bruising was based on spectral information. The model input information included full-wave information and valid information selected by the GA. The PLS-DA model is shown in Table II. Five misjudgments occurred when the full band information was used as input, with a classification accuracy of 92.86%. After the GA picked the variables, the number of false positives in the 70 prediction sets was zero with a 100% classification accuracy rate. ELM is a new machine learning algorithm based on a single-hidden layer neural network. The input weights and offset numbers are randomly selected. The initial value of ELM hidden layer neurons is 10, and the step size is set to 1, which gradually increases to the number of modeling samples. In the training, the sigmoidal, sine, and hardlim functions were used as excitation functions of neurons, and the number of spectra was used as the input. Table II shows the modeling results of the three excitation functions. As can be seen from Table II, when sine was the excitation function, the classification accuracy was the best. When the whole spectral data was used as the input, only three samples had misjudgment. When the data after variable selection was used as the input, only two data points had misjudgment, and the classification accuracy was 97.14%.

The discriminant results of the PLS model are plotted in Figure 12, with a model set RMSEC of 0.1306, a modeled set correlation coefficient of 0.928, a modeled set RMSEP of 0.1357, and a predicted set correlation coefficient (

) of 0.928. The results for the PLS model is shown in Figure 12.

FIGURE 12: Graph of spectral feature modelling results.

Figure 13 shows the predicted regression coefficients at different times after the bruising of the caper. The greater the weight of the spectral variables in the quantitative PLS model, the greater the regression coefficients. The greater the regression coefficient corresponding to the spectral variable, the more significant the contribution of the sample information to the model prediction.

FIGURE 13: Regression coefficient curves for the PLS-DA model.

This study demonstrated the high potential of hyperspectral imaging coupled with multivariate analysis to determine honey peach bruising. The PLS-DA regression model and the ELM regression model were used to establish the honey peach microjudgment discrimination model. In this study, based on the analysis of internal spectral features and external color features of bruised honey peach, the PLS-DA algorithm and ELM algorithm were used to establish the peach’s slight bruising discrimination model. The highest classification accuracy of color features was only 74.29%. In the spectral information analysis of the honey peach collision, the GA was used to virtually screen 176 spectral features. The PLS-DA model established based on the spectral all-band information was superior to the ELM model. After the GA selected 40 useful variables, there was no wrong number of samples in 70 prediction sets, and the classification accuracy was 100%. RMSEP was 0.1357, and the 

 was 0.928. The spectral information and image information were used as model inputs separately, and the model built by the spectral information was better than the image information. The reason may be because the softening and browning of internal tissues are not obvious after bruising of yellow peaches, and it is difficult to distinguish them in terms of appearance and color. The high accuracy of spectral information discrimination may be that the changes in thermal tissue characteristics caused by internal defects and physiological disorders are expressed through internal spectra. According to the spectral information, when the honey peach was slightly injured, the honey peach can still achieve its highest classification accuracy. In particular, online sensors will also be developed for honey peach sorting or post-harvest quality processing of other fruits.

The research was funded by the National Natural Science Foundation of China (No. 31760344), the Science and Technology Research Project of Education Department of Jiangxi Province (Grant No. GJJ200652), and the Science and Technology Research Project of the Education Department of Jiangxi Province (Grant No. 190306).

Conceptualization was done by Xiong Li; the data curation was performed by Xiong Li and Guantian Wang. Funding acquisition for the study was handled by Xiong Li and Yande Liu; Xiong Li and Yunjuan Yan handled the writing, editing, and review of the manuscript.

(6) G. ElMasry, C. Vigneault, and J. Qiao, 

(7) J.P. Cruz-Tirado, J. Pierna, and H. Rogez, 

(8) S.H. Park, Y.K. Hong, and S.B. Mubarakat, 

(9) K.B. Walsh, J. Blasco, and M. Zude-Sasse, 

By extracting the RGB, HSI, and grayscale information from a spectral range of 400–1100 nm and comparing the spectral features of sound and bruised peaches, the authors provided a classification system and theoretical basis for online fruit bruise detection.

Detection of Early Bruises in Honey Peaches Using Shortwave Infrared Hyperspectral Imaging

The feasibility of quantifying the soluble solids content of intact apples was investigated by visible and near infrared (vis–NIR) transmittance spectroscopy combined with the least squares support vector machines (LS-SVM) method. The spectra were pretreated by Savitzky-Golay smoothing, first and second derivatives, standard normal variate transformation, and multiplicative scatter correction. The regression models were developed by LS-SVM and partial least squares (PLS). The accuracy of the LS-SVM and PLS models was compared. 

The feasibility of quantifying the soluble solids content of intact apples was investigated by visible and near infrared (vis–NIR) transmittance spectroscopy combined with the least squares support vector machines (LS-SVM) method. The spectra were pretreated by Savitzky-Golay smoothing, first and second derivatives, standard normal variate transformation, and multiplicative scatter correction.

Yande Liu

Articles

Detection of the Early Fungal Infection of Citrus by Fourier Transform Near-Infrared Spectra

Detection of Early Bruises in Honey Peaches Using Shortwave Infrared Hyperspectral Imaging

Quantification of the Soluble Solids Content of Intact Apples by Vis–NIR Transmittance Spectroscopy and the LS-SVM Method