Application of hyperspectral technology in detection of agricultural products and food: A Review

Abstract Food is the foundation of human survival. With the development and progress of society, people increasingly focus on the problems of food quality and safety, which is closely related to human's health. Thus, the whole industrial chain from farmland to dining table need to be strictly controlled. Traditional detection methods are time‐consuming, laborious, and destructive. In recent years, hyperspectral technology has been more and more applied to food safety and quality detection, because the technology can achieve rapid and nondestructive detection of food, and the requirement to experimental condition is low; operability is strong. In this paper, hyperspectral imaging technology was briefly introduced, and its application in agricultural products and food detection in recent years was systematically summarized, and the key points in the research process were deeply discussed. This work lays a solid foundation for the peers to the following in‐depth research and application of this technology.

the market contained a strong carcinogenic "malachite green," which shocked the world and caused panic.
As the saying goes, "food is the most important thing for the people," which also contains a meaning that food safety and quality is greater than heaven, because it is closely related to the health of human. The emergence of food safety incidents poses new challenge to traditional detection methods of food quality and safety, including the detection of the whole industrial chain from farmland to dining table. Traditional methods have great limitations, such as long detection cycle, strong destructive, complex operation, and single-point detection. Consequently, spectral technology has gradually been introduced into food quality and safety detection, such as hyperspectral technology (Dale et al., 2013), infrared spectrum technology (Fu & Ying, 2016;Shi et al., 2012), and Raman spectrum technology (Ai et al., 2018).
In this paper, the principle and analysis process of hyperspectral imaging technology were briefly introduced, and the application of this technology in the field of agricultural products detection in recent years, including grains, fruits, vegetables, and meats, was summarized systematically. In addition, the deficiencies and key points in the research were discussed in depth. This work lays a solid foundation for the peers to the following in-depth research and application of this technology.

| HYPER S PEC TR AL IMAG ING TECHNOLOGY
As an emerging, nondestructive, and advanced optical technology, it is an image data technology with many narrow bands. It combines mechanical vision with spectral technology to detect the two-dimensional spatial and one-dimensional spectral information of the targets; thus, high-resolution image and spectral data are obtained.
Therefore, the emergence of hyperspectral technology makes it easier to detect objects that cannot be detected with wide band.
Moreover, compared with other optical technologies, hyperspectral image is closer to the real properties of objects. At present, this technology has developed rapidly, which can be divided into reflection imaging (Nicolaï, Lötze, Peirs, Scheerlinck, & Theron, 2006), fluorescence imaging (Vargas et al., 2010), and transmission imaging (Casasent, 2011). Among them, reflective imaging technology is the most commonly used. The hyperspectral imaging system is mainly consisted of four parts: hyperspectral camera, light source, carrier stage, and computer software and hardware (Figure 1a; Wu et al., 2012). The light emitted by the source is absorbed and then reflected by object surface ( Figure 1b; Li & Rao, 2011). After passing through the front lens and entrance slit, light with different wavelengths will have bend-divergence propagation of different level. Then, it converges at the collimation lens, light of different wavelengths form separate bands by splitting. Finally, the spectral signal will be presented to the detector through the imaging lens. The three-dimension data cube rich in image and spectral information are obtained by machine sweeping ( Figure 1c). Moreover, when choosing the light source, we should pay attention to highlight the object and weaken the background.
Meanwhile, to present useful signal as much as possible, the signalto-noise ratio of the image should be improved, thus reduce noise interference (Dong, Guo, Xu, & Xu, 2018). Imaging spectrometer is also called hyperspectral camera, which can absorb, process, and transmit the reflection spectrum of target, is one of the most core part of the whole hyperspectral system. The main function of the electronic control platform is to control the moving speed of the object and make it consistent with the sampling frequency and exposure time of the camera, thus prevent the phenomenon of missing or repeated acquisition. Data acquisition software mainly control the operation of relevant equipment through parameter setting, thus efficiently completing the data acquisition work.
The hyperspectral off-line data are processed by chemometrics and computer technology, which is mainly implemented in MATLAB and ENVI software. The general flow chart of data analysis is shown in Figure 2, in which preprocessing, variable selection, and modeling methods are the key steps in the whole analysis process, they all involve a variety of processing algorithms. The selection of algorithms has an important influence on the model accuracy and prediction performance of different variables. The purpose of pretreatment is to remove the noise fluctuation and baseline change generated in the process of data acquisition, so as to enhance the spectral signal. The commonly used spectral pretreatment methods include multiplicative scatter correction (MSC) and standard normal variate (SNV). The spectral information generated in data acquisition originates in the overlap of signals of various chemical substances of sample. The characteristic wavebands closely related to the variables are selected by some methods, which is helpful to improve the predictive effect of the model on  Table 1.

| APPLIC ATION OF HYPER S PEC TR AL TECHNOLOGY IN THE FIELD OF AG RI CULTUR AL PRODUC TS AND FOOD DE TEC TI ON
Food detection includes safety and comprehensive detection (Huang, Yao, Hui, Sun, & Xing, 2012). Among them, safety detection refers to the detection of substances that may cause harm to human health in food. Comprehensive detection is also divided into external and internal detection, that is, the detection of external defects and internal quality of objects. Detection is usually carried out by random sampling combined with chemical analysis, which will inevitably prolong the detection time and reduce the accuracy of the result. With the introduction of hyperspectral technology, it has been widely used in the field of food detection (Table 2).

| Plant-product industry
Planting industry is the basic sector of the whole agriculture and is also the foundation of human existence, such as grain, cotton, and oil.
Plant-product industry mainly cultivates a variety of crops, according to the property and purpose of products, which can be divided into grain crops, cash crops, feed and manure crops, vegetables, etc.
Grain crop is one of the most important food for human beings, but it is easily infected by fungi during growth and storage, which leads to the decline of the yield and nutritional value (Orina, Manley, & Williams, 2017). In order to identify the situation of food infected by fungi, we can only wait for visible colonies on the surface, or conduct early identification by microbial culture, but this method is time-consuming and laborious. Williams et al. (Williams, Geladi, Britz, & Manley, 2012) used hyperspectral to detect the changes of fungi on maize surface after infection with Fusarium verticillioides. It was found that the fungal change could be identified by hyperspectral technology in the early stage of infection; meanwhile, the content changed significantly after starch and protein were utilized by fungi.
However, the infection ability of fungi to different biological samples is different, which will interfere with the establishment of the model, so more systematic research is needed to verify its feasibility.
In addition, whether brown rice is infected by fungi during storage can also be detected by hyperspectral technology, and the spectral signal decreases with the increase of fungal colonies (Siripatrawan & Makino, 2015). Once grains are infected by fungi, these fungi usually produce toxins that can cause serious harm to human health, such as aflatoxin in corn (Fiore et al., 2010). At present, traditional methods cannot effectively identify aflatoxin in early stage, yet hyperspectral technology can quickly identify it within 48 hr after artificial inoculation with Aspergillus flavus, which may be related to the detection limit of the method. In the early stage, the amount of aflatoxin produced by fungi is low, and the consumption of aflatoxin cannot be avoided by the solution transfer in the traditional detection process, so it cannot be effectively identified by this method. However, TA B L E 1 Statistical tables of commonly used spectral pretreatment, variable selection, and modeling methods

Modeling methods
Multiplicative scatter correction (MSC); normalized; standard normal variate (SNV); Savitzky-Golay 1st order (SVG-1) and 2nd order derivatives (SVG-2; Crichton et al., 2017) Successive projections algorithm (SPA); regression coefficient (Caporaso, Whitworth, Grebby, et al., 2018) Genetic synergy interval partial least square (GA-Si-PLS) algorithm (Ling et al., 2017) Autoscale  Competitive adaptive reweighted sampling (Tian et al., 2018) Partial least squares discrimination analysis (PLS-DA; Sun et al., 2017) De-trending (Caporaso, Whitworth, Grebby, et al., 2018) Weighted values (Qu et al., 2017) Partial Hyperspectral technology can also be used to identify the damage of vegetables, so that there is quality assurance before sale, which cannot only ensure the freshness of vegetables in the storage cycle, but also make consumers feel at ease to buy. Mushroom is easy to be damaged in the process of transportation. To solve this problem, Gowen et al. (2008) effectively detected the damage of white mushroom during transportation through hyperspectral technology, and the technology can be used for rapid identification the damage of white mushroom on the production line. Therefore, these studies show that hyperspectral technology can effectively achieve the rapid and nondestructive detection of internal and external quality of fruits and vegetables, so as to ensure the freshness and quality.

| Animal husbandry
Meat products are nutritious and delicious, and are very popular with consumers. However, some illegal trader sell unqualified meat for their own huge benefit, which seriously violates the rights and interests of consumers. Kamruzzaman, Elmasry, Sun, & Allen (2011) not only used hyperspectral technology to quickly discriminate the muscle of three different parts of lamb, but also established prediction models for the identification of muscle pH, color, and mass loss (Kamruzzaman, Elmasry, Sun, & Allen, 2012). In order to further reveal the adulteration of meat products, the team (Kamruzzaman, Sun, Elmasry, & Allen, 2013) again used hyperspectral to quickly and effectively detect the content distribution of adulterated ingredients in lamb meat. Chicken is a kind of meat food which is easy to deteriorate even in low temperature. Total volatile basic nitrogen (TVB-N) is an important indicator to detect the deterioration of chicken. Hyperspectral imaging technology can detect it quickly and nondestructively (Khulal, Zhao, Hu, & Chen, 2016), prevent spoiled chicken from entering the market, and endanger human health.
Deep-frozen can prolong the shelf life of meat, which is a common method for meat storage. However, due to oxidation reaction and mechanical damage, the flavor of meat will decrease seriously during freezing. With carbonyl content as an indicator, the oxidative damage of pork myofibrils during frozen storage can be well measured by hyperspectral technology (Cheng, Sun, Pu, & Wei, 2018). With the pH value, color, and tenderness of beef as indicators, hyperspectral technology can detect the difference in beef with different freshness (Crichton et al., 2017), the water holding capacity of fresh beef (Eimasry, Sun, & Allen, 2012) and accurately classify the beef grade ). In addition, the content of melamine

| Aquatic farming industry
Parasite is major factor that endangers the health and growth of ani-  (2012) proposed an effective and rapid method to identify nematodes in cod slices, which proved the advantage of hyperspectral detection in meat parasites.
This method is suitable for large-scale detection of aquaculture industry. Meanwhile, hyperspectral technology has also been introduced into the detection of viable count on the surface of fish (Wu & Sun, 2013). For aquatic products, in addition to the detection of parasites and microorganisms, some chemical compounds are also important indicators of internal quality ).

Qu, Sun, Cheng, & Pu (2017) established a visual distribution map
of the moisture content of grass carp fillets during freeze-drying by hyperspectral technology. Moreover, they also successfully identified different grades of shrimp using hyperspectral technology (Qu et al., 2015). However, some important information may be lost in spectral preprocessing, and the established model is not universal.
Therefore, how to improve the accuracy and prediction performance of the model is one of the important research directions.

| Others
Hyperspectral technology has been widely used in the food field. In addition to the primary agricultural products and food mentioned above, this technology has also been applied to other kinds of food detection, such as nutriments (Shi et al., 2017) and cocoa beans (Caporaso, Whitworth, Fowler, & Fisk, 2018). Moreover, hyperspectral technology is first used to quantitatively predict the content of sucrose, caffeine, and triglycerides in single coffee bean (Caporaso, Whitworth, Grebby, & Fisk, 2018). The research of hyperspectral technology is not just a small-scale validation test, but also has been industrialized. The scientific and technological product line of Nongfu Mountain Spring (a famous drinking water manufacturing enterprise in China) is a typical example of industrialized application of hyperspectral imaging technology. Through hyperspectral photography system combined with computer technology, the types and area of fruit surface defects are identified, and the illumination sorting of fruit is realized. Meanwhile, the near-infrared detection system form spectral curve by irradiating fruit surface to realize nondestructive detection of sugar and acidity. Before doing this work, it is necessary to establish a very large database, the more databases there are, the higher the accuracy of detection data will be.

| D ISCUSS I ON
Hyperspectral imaging technology has unique advantages in the field of nondestructive testing compared to traditional methods.

| CON CLUS I ON AND PER S PEC TIVE
As a image data technology, hyperspectral imaging technology has the advantage of union of imagery and spectrum. It can simultaneously detect the surface and internal information of objects, so as to realize the rapid and nondestructive detection of food quality and safety. Therefore, it has been widely used in the field of food. In this paper, from the perspective of agricultural classification, the research progress of hyperspectral technology in primary agricultural products and food in recent years was systematically reviewed. In addition, the deficiencies and key points of this technology in the research were discussed in depth. This work lays a solid foundation for peers to quickly grasp the application progress of hyperspectral technology in the field of agricultural products and food, contributing to the in-depth research and application of this technology.
In China, solid-state brewing (Liu et al., 2004) has a very long history and peculiar culture, and the open fermentation of multi-strains is the main characteristics of making process. Baijiu and vinegar are typical representatives of solid-state brewing. Based on hyperspectral technology, the distribution of moisture and acidity in vinegar fermented grains is quickly detected (Zhu et al., 2016), enabling baijiu enterprises to find problems quickly, and adjust processes in time, thus ensure the product quality. It shows that it is feasible to apply hyperspectral imaging technology to the rapid detection of characterization indicators in solid-state fermentation process. As one of the six distilled spirits in the world (Zhao, Zheng, Song, Sun, & Tian, 2013), chinese baijiu is brewed by open fermentation condition with natural inoculation. Because of its complex fermentation system, it is difficult to effectively monitor the production process. With the arrival of the mechanization and intelligence of chinese baijiu, hyperspectral technology has a broad application prospect in the field of baijiu making, which will have important guiding significance for the transformation and upgrading of traditional technology and intelligent on-line monitoring of complex fermentation state.

ACK N OWLED G M ENT
This work was supported by Sichuan science and technology program (2019YJ0475).

CO N FLI C T O F I NTE R E S T
The authors declare that they have no conflicts of interest.

AUTH O R CO NTR I B UTI O N S
Min Zhu collected data, drafted the manuscript, and revised it critically. Dan Huang and Xin-Jun Hu revised the manuscript critically for important intellectual content. Wen-Hua Tong, Bao-Lin Han, and Jian-Ping Tian gathered data. Hui-Bo Luo contributed to the design of the work. All authors approve the final version of the manuscript and agree to be accountable for all aspects of the work.

E TH I C A L S TATEM ENTS
This study does not involve any human or animal testing.