Classification of Sugar Content of Kiwi Fruit Based on Deep Learning and Near Infrared Spectrum

doi:10.21203/rs.3.rs-3799363/v1

Download PDF

Research Article

Classification of Sugar Content of Kiwi Fruit Based on Deep Learning and Near Infrared Spectrum

https://doi.org/10.21203/rs.3.rs-3799363/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

In recent years, the kiwifruit industry has undergone rapid development. The quality of kiwifruit significantly impacts its market price. The current manual grading system for kiwifruit is inefficient and lacks accuracy. Therefore, the application of intelligent methods for the rapid classification of kiwifruit has become a prominent topic in kiwifruit research. Sugar content is a critical factor in evaluating kiwifruit quality. This study establishes a non-destructive detection and classification model for sugar content using near-infrared spectroscopy and deep learning techniques. Initially, near-infrared spectra and sugar content values of kiwifruit were gathered. Following the classification of kiwifruit samples based on sugar content standards, two preprocessing methods – MSC and SNV – along with four deep learning network structures – LeNet, AlexNet, VGG, and ResNet – were utilized to construct classification models for the spectra and sugar content. Through comparison of training set accuracy, test set accuracy, and confusion matrix of the test set, the combination of MSC preprocessing and the ResNet classification model was identified as the optimal approach, achieving a test set accuracy of 100%. These findings indicate that the classification model established using near-infrared spectrum and deep learning effectively promotes kiwifruit classification based on sugar content.

Near infrared spectrum

Deep learning

Kiwi fruit

Sugar content

Kiwifruit, known as the king of fruits, is originally from China and is cultivated in regions such as Shaanxi, Sichuan, Hunan, Guangdong and Guangxi. Kiwi fruit is rich in vitamins, dietary fiber, amino acids, organic acids, and other beneficial nutrients and active substances. It also has various health promoting properties such as antioxidant, anti-tumor, immune regulation, fatigue relief, weight loss, and constipation relief. [1].

With its abundant and sweet-tasting juice, kiwifruit has gained popularity among consumers. In recent years, both domestic and international production and sales of kiwifruit have been steadily increasing[2]. The sugar content of kiwifruit is a crucial factor influencing its marketability, and grading kiwifruit based on sugar content could contribute to the development of the kiwifruit market.

The sugar content of kiwifruit is calculated by converting the refractive index of kiwifruit juice to a specific concentration of sucrose solution[3]. Traditional methods for measuring kiwifruit sugar content are destructive, involving the extraction of kiwifruit juice followed by the use of a sugar content analyser. This approach not only requires substantial manual labour and is time-consuming but also involves sampling, which limits the ability to test each individual kiwifruit.

Near-infrared spectroscopy (NIR) is a widely utilized analytical technique in the field of fruit detection, offering rapid, accurate, and non-destructive characteristics. The wavelength distribution of near-infrared light ranges from 780 to 2530 nm, where molecular vibrations related to hydrogen-containing groups such as O-H, C-H, and N-H produce detectable overtones and combinations[4] [5]. By analyzing the position and intensity of the obtained near-infrared absorption spectra, the sugar content of kiwifruit can be determined. Compared with traditional machine learning algorithms, deep learning algorithms demonstrate enhanced feature extraction capabilities and better model generalization, effectively handling NIR data.

Deep learning classification networks possess robust modeling capabilities, addressing the issue of underfitting commonly observed in traditional machine learning algorithms when dealing with large sample sizes[6] [7]. Furthermore, they have automatic feature extraction function, eliminating the need for the step of extracting feature wavelengths and effectively capturing valuable spectral information[8].

Mishra et al. [9] provided a critical and comprehensive review of the main advantages and potential pitfalls of deep learning techniques for spectral data modeling, but did not analyze specific data using deep learning network models. Rong et al. [10] constructed a VIS-NIR spectral database containing multiple peach varieties and achieved multi-identification of peach varieties by building a one-dimensional convolutional neural network, but the deep learning network model they used is relatively simple and its accuracy is not enough. Chen et al. [11] identified cumin and fennel varieties based on the difference in near-infrared spectral signal intensity, combined with an improved multiscale fusion convolutional neural network and a near-infrared spectral-based bidirectional long short-term memory network. Abasi et al. [12] utilized wavelet transform for apple spectral processing and established a well-performing PLSR soluble solid prediction model through comparative analysis, but they did not integrate deep learning into their approach, relying solely on traditional machine learning methods. Cortés et al. [13] utilized diffuse reflectance in combination with partial least squares regression (PLSR) to build a prediction model, effectively extracting mango sample features using near-infrared spectral technology to meet the real-time monitoring demand for mango ripeness, albeit with a complex data preprocessing procedure. Based on existing research, it is evident that there is significant potential for kiwifruit classification detection using near-infrared spectroscopy combined with deep learning [14][15].

This study focuses on Xuxiang kiwifruit from Shaanxi as the research subject, collecting spectra of kiwifruits with different sugar contents. Upon processing the spectral data, various deep learning network models are employed to build NIR models, and the optimal sugar content classification model is obtained by comparing the accuracy of kiwifruit sugar content classification under different network models. This study achieves kiwifruit grading based on sugar content, providing a rapid and accurate method for kiwifruit quality classification.

2.1 Materials

The experimental samples used in this study were the same batch of Xuxiang kiwifruit purchased from the Bairui Kiwifruit Research Institute in Zhouzhi, Shanxi Province. The total number of samples was 140. Then all samples were uniformly placed in a freezer at a constant temperature of 4℃ for refrigeration[16].

2.2 Instruments and equipment

In this study, the portable spectrometer of NIL-R210 model from PYNECT and the accompanying ISC-NIRS software were used for spectral data acquisition.

The sugar content of the samples was measured using ATAGO's PAL-BXACID8 handheld sugar meter.

2.3 Data Collection

Before collecting spectral data from kiwifruit samples, a thin glass capable of holding objects was fixed 1cm above the light source of the spectrometer, ensuring a consistent distance between each kiwifruit sample and the light source during spectral acquisition for reference spectrum collection. The standard whiteboard was placed on the transparent glass surface and used to collect spectra, with the obtained spectrum from the whiteboard serving as the reference spectrum. After saving the reference spectrum, the spectral data of the kiwifruit samples were collected. Each kiwifruit sample was sequentially placed on the transparent glass, aligning the focal point of the light source with the equator of the kiwifruit. Subsequently, the spectral data of each kiwifruit sample were collected and saved as CSV files. The acquisition conditions included 10 scans with an integration time of 2.54ms. Each kiwifruit was scanned once at four different points along its equator, resulting in four sets of spectral data. The average of these four sets was considered as the final spectral data for each kiwifruit sample [17].

After completing the collection of spectral data, the kiwifruit samples were peeled and cut into pieces. The pieces were then placed into a juicer to extract the kiwifruit juice. The obtained kiwifruit juice was filtered using non-woven fabric to obtain clear kiwifruit juice. Subsequently, the juice was dropped into a handheld refractometer for measuring the sugar content, and the sugar content data was recorded.

2.4 Data Processing

After collecting the sugar content values of kiwifruit, a grading standard for kiwifruit sugar content was established. According to this standard, the kiwifruit samples were classified into different grades for establishing classification models.

During the modelling process, preprocessing of the original spectra is an important step in the quantitative and qualitative analysis of near-infrared spectra [18]. This is because the spectra obtained from the spectrometer may contain background information or noise in addition to useful samples. Therefore, it is necessary to preprocess the spectral data to establish accurate and stable calibration models. Standard Normal Variate correction (SNV) [19] and Multiplicative Scatter Correction (MSC) algorithm [20] were used to preprocess the collected spectral data to eliminate errors such as scattering during spectral acquisition, instrument background interference, and baseline drift caused by the light source.

The Kennard-Stone algorithm (KS) was used to divide the kiwifruit samples into training and testing sets [21][22]. The ratio of training set to testing set is 2:1. The principle of the KS algorithm is to treat all samples as candidate samples for the training set, and then select samples into the training set one by one. The specific steps of the KS algorithm are as follows:

Select the two samples with the furthest Euclidean distance to enter the training set;

Calculate the Euclidean distance from each remaining sample in the sample set to each known sample in the training set;

Compare and select the two samples with the maximum of minimum Euclidean distances to enter the training set;

Repeat the above steps until the number of samples in the training set reaches the standard.

The KS algorithm ensures that the samples in the training set are uniformly distributed according to spatial distance [23]. The formula for calculating Euclidean distance [24] is shown in Eq. (1).

$$\begin{array}{c}{d}_{x}\left(p,q\right)=\sqrt{\sum _{j=1}^{N}{\left[{x}_{p}\left(j\right)-{x}_{q}\left(j\right)\right]}^{2}} \left(1\right)\end{array}$$

Where, ${x}_{p}$ and ${x}_{q}$ represent two different samples, and N represents the number of spectral wave points of the sample.

2.5 Classification Model

In this study, four deep learning network models, namely LeNet [25], AlexNet [26], VGG [27], and ResNet [28], were employed to conduct classification training on the spectral data.The best performing model was selected by comparing the predictive performance indicators of four models. To enable the deep learning classification network to process one-dimensional kiwifruit spectral data, an additional dimension was added to the one-dimensional spectral data before entering the model for training to adapt it to the classification network.

Traditional deep learning classification networks consist of input layers, convolutional layers, and output layers. The LeNet network is an early convolutional neural network with a relatively simple structure, featuring only two convolutional layers. The AlexNet network deepened the network based on the LeNet network, performed data augmentation when processing data, and introduced the Dropout method to avoid overfitting. The VGG network, based on the AlexNet network, used three 3x3 convolution kernels to replace the 7x7 convolution kernel and two 3x3 convolution kernels to replace the 5x5 convolution kernel, thereby increasing the depth of the network and to some extent improving the effectiveness of the neural network. The ResNet network incorporated residual block structures before each convolutional layer, saving model training time while eliminating the problem of network degradation, which occurs as the number of network layers increases and leads to a decrease in the prediction rate of the training and testing sets.

The ResNet network's convolutional layers are divided into five layers. The first convolutional layer requires the use of the ReLU activation function and the MaxPool1d maximum pooling operation. The size of the convolutional kernel used in the first layer is 7, while the sizes of the convolutional kernels for the remaining four layers are 3. An AdaptiveAvgPool1d function is applied after the convolutional layers to connect to an average pooling layer. The structural layer parameters of the ResNet network are shown in Table 1.

Table 1

ResNet network parameters
Network layer	Model parameter
Input layer	Near infrared spectral data
conv1	kernel_size = 7, ReLU, MaxPool1d, output channels = 64
Max Pooling layer	kernel_size = 3
conv2	Number of residual blocks = 2,
	kernel_size = 3, output channels = 64
	kernel_size = 3, output channels = 64
conv3	Number of residual blocks = 2,
	kernel_size = 3, output channels = 128
	kernel_size = 3, output channel number = 128
conv4	Number of residual blocks = 2,
	kernel_size = 3, output channels = 256
	kernel_size = 3, output channels = 256
conv5	Number of residual blocks = 2,
	kernel_size = 3, output channels = 512
	kernel_size = 3, output channels = 512
Average pooling layer	AdaptiveAvgPool1d(1)
Fully connected layer	The eigenvectors of the above layers are flattened into one-dimensional vectors
Output layer	One neuron

2.6 Evaluation index

In this study, accuracy and confusion matrix were used as evaluation indicators for classification models.

Accuracy refers to the ratio between the number of correct samples predicted by the model and the number of all samples. The higher is the accuracy, the better is the classification performance of the model [29]. The calculation formula is as Eq. (2).

$$\begin{array}{c}Accuracy=\frac{N}{T}\end{array}\left(2\right)$$

Where, N indicates the number of samples that are predicted correctly, and T indicates the number of all samples.

A confusion matrix is a table that summarizes the prediction results of a classification model [30]. The rows of the matrix represent the true values, and the columns represent the predicted values. By examining the diagonal elements of the confusion matrix, we can determine the number of correctly classified samples for each class [31].

3.1 Original spectral analysis

The spectral data of kiwifruit samples collected by the spectrometer ranges from 900 to 1700nm, with a total of 160 characteristic bands. Due to changes in molecular stretching, vibration, and bending states, different energy levels result in the absorption of near-infrared spectra, typically exhibiting characteristic absorption peaks. The original absorbance spectra of 140 kiwifruit samples are shown in Fig. 1. From the collected near-infrared spectral data, it can be observed that the spectral trends of each sample are consistent, indicating that the spectral information of kiwifruit samples is basically uniform. Additionally, different samples exhibit varying absorbance, suggesting differences in the chemical composition of each sample. The collected near-infrared spectral data meets the criteria for establishing a classification model [32].

Due to changes in molecular stretching, vibration, bending states, and other conditions, near-infrared spectrum absorption occurs between different energy levels, typically manifesting as characteristic absorption peaks in near-infrared spectra [33]. The average spectrum of all kiwifruit samples was obtained by averaging the original spectra, as depicted in Fig. 2. The absorption peaks of the average kiwifruit spectrum can reflect the overall absorption peak characteristics of the kiwifruit samples. By employing mathematical methods, the absorption peaks of the kiwifruit absorbance spectrum were identified, and the corresponding near-infrared absorption bands for kiwifruit are presented in Table 2.

Table 2

Absorption bands of kiwi fruit
Absorption	Wave length/nm
C-H(First order frequency doubling)	1688
C-H(Chemical bond)	1460
C-H(Second order frequency doubling)	1187
O-H(Second order frequency doubling)	976

From Fig. 1, Fig. 2 and Table 2, it can be observed that kiwifruit exhibits significant absorption peaks around 976 nm, 1187 nm, 1460 nm, and 1688 nm. These absorption peaks correspond to the functional groups of second overtone C-H, C-H bond, first overtone C-H, and second overtone O-H [34]. These functional groups are associated with the sucrose content in kiwifruit. Therefore, based on the collected kiwifruit spectra, they can be utilized for sugar level classification and detection.

3.2 Establishment of kiwi sugar degree model

3.2.1 Sugar content data processing and spectral preprocessing

According to the sugar content classification criteria, the 140 collected kiwifruit samples were divided into three categories: low sweetness (designated as Class 0, with sugar content below 11% BRIX), medium sweetness (designated as Class 1, with sugar content between 11% and 13% BRIX), and high sweetness (designated as Class 2, with sugar content above 13% BRIX). According to this classification standard, the number of kiwifruit samples for each sweetness level is summarized in Table 3 as follows:

Table 3

Number of kiwi fruit by sugar grade
Sugar degree	Class 0(Low sweetness)	Class 1(Medium sweetness)	Class 2(High sweetness)
Number	52	61	27

The original spectrum is affected by scattering, instrumental background interference, optical path length, and light drift [35]. Therefore, preprocessing of the original spectrum is necessary. The SNV algorithm and MSC algorithm are used for preprocessing the collected spectral data. The MSC algorithm is employed to eliminate the interference caused by multi-scattering effects in the samples, with the primary goal of reducing the impact of scattering effects on the spectral signal [36]. Meanwhile, the SNV algorithm is used to correct the influence of scattering and path length changes on NIR diffuse reflectance spectra [37]. The processed spectrum is shown in Fig. 3, where it can be observed that the boundaries of the peaks and valleys in the original spectrum are not clear. However, after preprocessing, the spectral peaks and valleys become more distinct.

3.2.2 Model training and prediction

The untreated training set kiwifruit data is imported into the classification network, and the training continues until the loss function reaches its minimum and stabilizes. At this point, the training phase of the network model is complete, and the accuracy and loss functions of the training set are shown in Fig. 4. It is evident that the ResNet network requires the fewest training iterations and achieves the highest accuracy.

In order to identify the best performing pre-processing method and develop a classification model for kiwifruit quality grading based on sugar content, we will utilize different pre-processing methods on the spectral data and train them using four deep learning models: LeNet, AlexNet, VGG, and ResNet. The accuracy of each pre-processed dataset using different models on the training and testing sets is shown in Table 3, with 1000 iterations conducted for each experiment. We will select the best performing pre-processing method and deep learning model based on the results in Table 4.

Table 4

Modelling result
	Training set($\text{A}\text{c}\text{c}\text{u}\text{r}\text{a}\text{c}\text{y}$)	Test set($\text{A}\text{c}\text{c}\text{u}\text{r}\text{a}\text{c}\text{y}$)
Basic data + LeNet	0.8351	0.3918
SNV + LeNet	0.9488	0.4048
MSC + LeNet	0.9597	0.5000
Basic data + AlexNet	0.9186	0.5238
SNV + AlexNet	0.9319	0.6429
MSC + AlexNet	0.9143	0.8810
Basic data + VGG	0.9762	0.6429
SNV + VGG	0.9808	0.5952
MSC + VGG	1.0000	0.9762
Basic data + ResNet	0.8969	0.9048
SNV + ResNet	0.9381	0.9762
MSC + ResNet	1.0000	1.0000

According to Table 4, it can be observed that using data pre-processing methods yields better results compared to not using any pre-processing. The MSC method outperforms the SNV algorithm and further improves the prediction accuracy of the model. After applying the MSC pre-processing method to the spectral data, different models were trained, and the confusion matrix of the testing set is shown in Fig. 5. The LeNet network incorrectly identifies class 0 kiwifruits as class 1. AlexNet correctly classifies most of the kiwifruits but misclassifies class 0 and class 2 as class 1. VGG network achieves mostly accurate predictions, but there are instances where it misclassifies class 0 as class 1. On the other hand, ResNet successfully classifies all the kiwifruits accurately. Therefore, ResNet demonstrates the best classification performance among the tested models.

The theoretical optimal combination for the kiwifruit sugar content classification model is MSC pre-processing combined with the ResNet model. The modeling approach is the primary factor influencing model accuracy, while the pre-processing method is a secondary factor. Comparing the modeling approaches, it can be concluded that the ResNet network performs the best, followed by the VGG network, then the AlexNet network, and finally the LeNet network with the lowest performance.

This study collected near-infrared spectroscopy data and sugar content values of kiwifruit samples from the same origin and variety using a spectrometer and a sugar content meter. The near-infrared spectroscopy data was pre-processed using the MSC and SNV algorithms to eliminate the influence of noise, scattering, and other factors. Four kiwifruit sugar content classification models, LeNet, AlexNet, VGG, and ResNet, were established. The conclusion drawn was that under the same training conditions, the model utilizing the MSC pre-processing algorithm and the ResNet network achieved the highest prediction accuracy on the test set, reaching 100%. The research indicates that the combination of near-infrared spectroscopy technology and deep learning techniques is feasible for kiwifruit sugar content classification detection, and the established model can accurately classify kiwifruit of different sugar content levels.

Based on the autonomously collected kiwifruit spectroscopy and sugar content dataset, this study successfully achieved the classification detection of kiwifruit sugar content levels. Building upon this research, further in-depth investigations can be pursued in the following two aspects: (1) When establishing prediction models, the samples used belong to the same variety and originate from the same location, collected within the same time frame. The predictive effectiveness for kiwifruit grown in other locations and of different varieties remains unknown. Future research can involve gathering kiwifruit samples from different locations and varieties to establish and experiment with a universal model for sugar content level detection. (2) Kiwifruit quality encompasses acidity, hardness, vitamins, amino acids, and more. Subsequent studies can involve collecting samples with multiple physicochemical properties to achieve simultaneous multi-index detection of kiwifruit, thereby increasing the diversity of kiwifruit grading and further enhancing the economic benefits of kiwifruit sorting.

Author Contribution

Liu. and Wang. wrote the main manuscript text.All authors reviewed the manuscript.

Research progress on nutrient composition, biological activity and comprehensive utilization of kiwifruit. Lu Yu Dan; Liu Xiaochi; Feng Xin; Gao Minxia; Larry Union; Chen Guixin; Chen Yiting. Southeast Horticulture,2022
Guo Linlin P, Rongli W (2022) Kiwi fruit nutritional quality comprehensive evaluation [J]. J fruit trees 33(10):6:1864–1872 The DOI: 10.13925 / j.carol carroll nki GSXB. 20220289
Juan P, Xue-Guang S Chu Gang-hui. Rapid detection of miniature near-infrared spectroscopy buy mention apricot sugar [J]. Journal of analytical laboratory, and 2023 (10): 1285–1291. The DOI: 10.13595 / j.carol carroll nki issn1000-0720.2022.081004.
Liu Cuiling W, Jingzhu Sun Xiaorong. Research on Near infrared Spectroscopy in Food Quality Detection [M].: China Machine Press,201511
Martín GF J. A Guide to Near-Infrared Spectroscopy[M].Nova Science Publishers, Inc.:2023-08-21
Ciurczak WE, Igne B, Workman J et al Handbook of Near-Infrared Analysis, Fourth Edition[M].CRC Press:2021-01-12
Zhou Feiyan J, Linpeng D Jun. Review of convolutional neural networks [J]. Chin J Computers 2017, 40(06):1229–1251
Zhijun S, Lei X, Xu Y et al Application Res Computers 2012, 29(08):2806–2810. (in Chinese)
Mishra P, Passos D, Marini F et al (2022) Deep learning for near-infrared spectral data modelling: Hypes and benefits[J]. TRAC Trends Anal Chem, : 116804
Rong D, Wang H, Ying Y et al (2020) Peach variety detection using VIS-NIR spectroscopy and deep learning[J]. Comput Electron Agric 175:105553
Chen Y, Bin J, Zou C et al (2021) Discrimination of fresh tobacco leaves with different maturity levels by near-infrared (NIR) spectroscopy and deep learning[J]. Journal of Analytical Methods in Chemistry, 2021
Abasi SM, SaeidJamshidi BF, Hadi DKM (2019) Rapid measurement of apple quality parameters using wavelet de-noising transform with Vis/NIR analysis[J]. Scientia horticulturae, p 252
Cortés V, Ortiz C, Aleixos N et al (2016) A new internal quality index for mango and its prediction by external visible and near-infrared reflection spectroscopy[J]. Postharvest Biol Technol 118:148–158
Yu X, Lu H, Wu D (2018) Development of deep learning method for predicting firmness and soluble solid content of postharvest Korla fragrant pear using Vis/NIR hyperspectral reflectance imaging[J]. Postharvest Biol Technol 141:39–49
Sa I, Lim JY, Ahn HS et al (2022) deepNIR: Datasets for generating synthetic NIR images and improved fruit detection system using deep learning techniques[J]. Sensors 22(13):4721
Chen C, Yang B, Si R et al (2021) Fast detection of cumin and fennel using NIR spectroscopy combined with deep learning algorithms[J]. Optik 242:167080
Beć KB, Grabska J, Huck CW (2021) Principles and applications of miniaturized near-infrared (NIR) spectrometers[J]. Chemistry–A Eur J 27(5):1514–1532
Yang J, Wang J, Lu G et al (2021) TeaNet: Deep learning on Near-Infrared Spectroscopy (NIR) data for the assurance of tea quality[J], vol 190. Computers and Electronics in Agriculture, p 106431
Tang Ruohan L, Xiuhua L, Xuegang et al (2019) Detection of fiber content in sugarcane by transmission vision-near-infrared spectroscopy [J]. Spectrosc Spectr Anal 43(08):2419–2425
Zhan Ke C, Jiwang, Xu Y et al Rapid detection method of freshness of frozen crayfish based on near-infrared spectroscopy [J/OL]. Food Sci : 1–13 [2023-10-15].
Dongjie L, Meng W, Xiaoshuan Z et al Fresh grape sugar content detection system based on Vis/NIR spectral sensing [J]. Spectrosc Spectr Anal 2023, 43(07):2146–2152
Yuling WANG, Jingru W Liu Hong Research progress of near infrared spectroscopy in quality detection of edible fungi [J/OL]. Food Ind Sci Technol : 1–15 [2023-10-15].
Li Xiaoli Z, Dongyi D, Yulun et al Rapid spectral detection method of main inclusions in fresh tea leaves based on Convolutional neural network [J]. J China Agricultural Univ, 201,26(11):113–122
Jinpeng SONG, Kun L et al (2023) (7) Zhang Chi. Study on classification of bunt wheat kernel based on deep learning and vision-near-infrared spectroscopy [J]. Journal of analysis and test, and : 784–793. The DOI: 10.19969 / j.f XCSXB. 23030501
LeCun Y, Bottou L, Bengio Y, Haffner P (1998) Gradient-based learning applied to document recognition.Proceedings of the IEEE. 86(11): 2278–2324
Krizhevsky A, Sutskever I, Hinton GE (2012) Imagenet classification with deep convolutional neural networks[C]//Advances in neural information processing systems. : 1097–1105
Simonyan K, Zisserman A (2014) Very deep convolutional networks for large-scale image recognition[J]. arXiv preprint arXiv:1409.1556,
He K, Zhang X, Ren S et al (2016) Deep residual learning for image recognition[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. : 770–778
Sun T, Lin H, Xu H et al (2009) Effect of fruit moving speed on predicting soluble solids content of 'Cuiguan' pears (Pomaceae pyrifolia Nakai cv. Cuiguan) using PLSR and LS-SVM regression [J], vol 51. Postharvest Biology & Technology, pp 86–90. 1
A review of the application of near-infrared spectroscopy (NIRS) in forestry. Wang Yizhi;Xiang Jianying;Tang Yu;Chen Wanjun;Xu Yuanjie.Applied Spectroscopy Reviews,2022
Jiang Xiaogang Z, Mingwang Y, Jinliang et al (2023) Parameter optimization of apple sugar content model based on near-infrared online device [J]. Spectrosc Spectr Anal 43(01):116–1211
KONG HJ, ZHAO GL, Wang WL et al (2015) Identification and comparison of bonded aroma components in six fruits by gas chromatography-mass spectrometry _. Kong H J Food Science 36(14):140–144
Simin H, Yufei H, Yanlong C et al Determination of pesticide residues in citrus by magnetic covalent organic skeleton solid phase extraction with high performance liquid chromatography [J]. Anal Chem, 20, 48 (10):1392–1399
ZHANG Y (2021) On-line detection of peach quality during storage based on visible/near infrared spectroscopy [D]. East China jiaotong university, DOI: 10.27147 /, dc nki. Ghdju. 2021.000266
Li Ming (2018) Research on optimization method of visible/near-infrared spectrum prediction model for fruit quality [D]. China Agricultural University
Sheng G Wang Qiaohua. Non-destructive detection of red sugar content and water content based on visible/near-infrared transmission spectroscopy [J]. Chin Opt, 201, 14(03):566–577
Le BT (2020) Application of deep learning and near infrared spectroscopy in cereal analysis[J]. Vib Spectrosc 106:103009

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Classification of Sugar Content of Kiwi Fruit Based on Deep Learning and Near Infrared Spectrum

Status:

Version 1

Abstract

Figures

1. Introduction

2 Sugar classification principle

2.1 Materials

2.2 Instruments and equipment

2.3 Data Collection

2.4 Data Processing

2.5 Classification Model

2.6 Evaluation index

3 Modelling and experimental analysis

3.1 Original spectral analysis

3.2 Establishment of kiwi sugar degree model

3.2.1 Sugar content data processing and spectral preprocessing

3.2.2 Model training and prediction

4 Conclusions

Declarations

Author Contribution

References

Additional Declarations

Status:

Version 1