Elsevier

Biosystems Engineering

Volume 166, February 2018, Pages 150-160
Biosystems Engineering

Research Paper
Utilisation of visible/near-infrared hyperspectral images to classify aflatoxin B1 contaminated maize kernels

https://doi.org/10.1016/j.biosystemseng.2017.11.018Get rights and content

Highlights

  • Visible/NIR hyperspectral imaging method for detecting pure AFB1 contamination.

  • Classification for maize kernels of different varieties is presented.

  • Detection of AFB1 coated on maize kernel surface possible at as low as 10 ppb.

  • Factorial discriminant analysis revealed important wavelengths for classification.

  • Variations in maize variety have no influence on AFB1 contamination classification.

A visible/near-infrared (VNIR) hyperspectral imaging (HSI) system (400–1000 nm) was used to assess the feasibility of detecting aflatoxin B1 (AFB1) on surfaces of 600 kernels of four maize varieties from different regions of the U.S.A. i.e. Georgia, Illinois, Indiana and Nebraska. For each variety, four AFB1 solutions (10, 20, 100 and 500 ppb) were artificially applied on kernel surfaces. Similarly, a control group was generated from 30 kernels of each variety treated with a solution of methanol. Principal component analysis (PCA) was used to reduce dimensionality of the HSI data followed by the application of factorial discriminant analysis (FDA) on the principal component variables. PCA results showed a pattern of separation between uncontaminated and contaminated kernels for all varieties except for Indiana and pooled samples. FDA showed the ability to predict AFB1 contamination of each variety with over 96% validation accuracy while prediction for AFB1 contamination group membership of pooled samples reached 98% accuracy in validation. Variation in the spectra of AFB1 contaminated kernels could have caused the variation in the predicted AFB1 contamination group membership. The PCA and FDA models where influenced by the chemical information from Csingle bondH, Nsingle bondH and Osingle bondH bonds of VNIR spectral regions. This study presents the potential of using VNIR hyperspectral imaging in direct AFB1 contamination classification studies of maize kernels of different varieties. The study further suggests that varietal differences of maize kernels may have no influence on AFB1 contamination classification.

Introduction

Maize is one of the major agricultural cereals grown worldwide. However, maize is susceptible to infection by toxigenic fungi, both in the field and during post-harvest storage. The fungal infections can reduce the quality and nutritive value of grains due to contamination with fungal secondary metabolites such as mycotoxins, which are toxic to both livestock and humans (Wu & Sun, 2013a). Aflatoxins are mycotoxins produced by Aspergillus species of fungi, such as Aspergillus flavus and Aspergillus parasiticus (Pearson et al., 2004, Shahin and Symons, 2011). Aflatoxin B1 (AFB1) is the most potent carcinogen and has been directly linked to long term adverse health effects, such as lung cancer in man and liver cancer in some animal species (Abrar et al., 2013). Consequently, around the world, organisations such as the International Agency for Research on Cancer (IARC) have recognised AFB1 as a Group 1 carcinogen for animals and humans (Fernández-Ibañez, Soldado, Martínez-Fernández, & De la Roza-Delgado, 2009). At the same time, governments such as U.S.A., through the U.S. Food and Drug Administration, have long been strictly monitoring and regulating aflatoxin levels in human food and animal feed, and have set 20 ppb as the concentration limit for maize (Abbas et al., 2006, Pearson et al., 2004). Therefore, the early detection of fungal infection and mycotoxin contamination of maize plants and kernels can be useful to prevent the entry of mycotoxins into the food chain (Del Fiore et al., 2010, Williams et al., 2012).

Commonly used analytical chemical methods to detect the toxicity in grains and feeds include enzyme-linked immunosorbent assay (ELISA), molecular identification approaches and chromatography techniques such as thin layer, gas and high performance liquid chromatography (HPLC). These methods have merit for being accurate and selective with very low detection limits. However, they are also expensive, destructive, and can be difficult and time-consuming. Hence, they are not well suited for handling a large number of grain samples. In order to detect mycotoxins on an industrial scale, it is critical and necessary to develop a rapid and non-destructive method. A faster, non-destructive method for detecting fungal and mycotoxin contamination is offered by spectroscopy. In this regard, spectroscopy related studies have been carried out, for example, Dowell, Ram, and Seitz (1999) used near infrared (NIR) spectroscopy to detect scab damage, vomitoxin and ergosterol in wheat. They showed that all scab-damaged kernels were correctly identified and more kernels with vomitoxin were identified by their method than did a visual inspection. In addition, vomitoxin and ergosterol were predicted with standard errors of 40 and 100 ppm, respectively. Pearson, Wicklow, Maghirang, Xie, and Dowell (2001) detected aflatoxin in maize with transmittance and reflectance spectroscopy. Using discriminant analysis and partial least squares regression, the authors analysed transmittance spectra (500–950 nm) and reflectance spectra (550–1700 nm) obtained on whole corn kernels that exhibited various levels of bright greenish-yellow fluorescence and reported that over 95% of the kernels were correctly classified as containing either high (>100 ppb) or low (<10 ppb) levels of aflatoxin. Similar results were obtained when using either transmittance or reflectance and when using either discriminant analysis or partial least squares regression. Dowell, Pearson, Maghirang, Xie, and Wicklow (2002) detected fumonisin in maize kernels infected with Fusarium verticillioides using reflectance and transmittance visible and near infrared spectroscopy. The authors reported that kernels with >100 ppm and <10 ppm were correctly classified as fumonisin positive or negative respectively, and models based on reflectance spectra had higher correct classification than models based on transmittance spectra. Wang, Dowell, Ram, and Schapaugh (2004) used a diode-array reflectance NIR spectrometer (400–1700 nm) to classify healthy and fungal-damaged soybean kernels and discriminate among various types of fungal damage. Partial least squares (PLS) and neural network models were developed by the authors and the highest classification accuracy was over 99% when spectra of 490–1690 nm were used with a two-class PLS model, whereas a five-class classification with neural networks yielded higher classification accuracy than PLS model. Berardo et al. (2005) detected kernel rots and mycotoxins in naturally and artificially contaminated maize samples using NIR reflectance spectroscopy and the samples were analysed for fungal infection, ergosterol and fumonisin B1 content. The authors indicated that the best predictive ability for global fungal infection and F. verticillioides was produced with a calibration model utilising maize kernels and maize meals, respectively. However, the problem with conventional spectroscopy is that it provides an average spectrum of a particular sample without any spatial information (Mishra et al., 2015, Wang et al., 2015a), meaning that trace mycotoxins in a large sample may be missed (Polder, Van Der Heijden, Waalwijk, & Young, 2005).

Hyperspectral imaging (HSI) is an emerging technology that integrates both imaging and spectroscopy into one system capable of recording both spatial and spectral properties of a given sample (Ropodi et al., 2016, Shrestha et al., 2016, Siche et al., 2016, Wu and Sun, 2013b). Because HSI obtains a complete spectrum for each pixel in an image (Wu & Sun, 2013b), it is possible to identify unique spectral signatures at multiple locations within a given sample. These signatures are the result of the physical and chemical characteristics of the particular material analysed (Shrestha et al., 2016). The information from VNIR hyperspectral imaging has been used in varietal identification studies for maize kernels (Wang et al., 2016, Zhang et al., 2012) using principal component analysis (PCA) and factorial discriminant analysis (FDA) based on chemometric techniques. Furthermore, investigations involving VNIR hyperspectral imaging or short-wave NIR hyperspectral imaging have shown the feasibility of aflatoxin detection and fungal identification studies on maize (Pearson and Wicklow, 2006, Wang et al., 2015a, Wang et al., 2015b, Williams et al., 2012). However, these studies involved kernels from a single maize variety, thus requiring further investigation to examine the influence of the variation in kernels due to maize variety on aflatoxin or fungal detection.

NIR spectroscopy measures overtones and combination bands of the fundamental molecular vibrations found in the infrared region associated mainly with Csingle bondH, Nsingle bondH, and Osingle bondH functionalities which are due to hydrogenic stretching, bending, or deformation vibrations (Berardo et al., 2005, Shenk et al., 2008).

Stretching vibrations occur at shorter wavelengths such as those used in this study (400–1000 nm). A sample material such as maize kernel or aflatoxin in this case, selectively absorbs NIR radiation that yields information about the molecular bonds within the material being measured. Shenk et al. (2008) stated that an NIR absorption band is produced when NIR radiation at a specific frequency (wavelength) vibrates at the same frequency (wavelength) as a molecular bond in the sample. Therefore, various functional groups at specific spectral bands may be correlated to the major maize kernel or aflatoxin constituents. According to Fernández-Ibañez et al. (2009) and Del Fiore et al. (2010), the spectral information in the visible region of between 400 and 600 nm corresponds to colour changes especially of carotenoids and chlorophyll in cereal grains. Several infrared spectroscopic methods have been developed to identify aflatoxin and fungi in cereal grains (Fernández-Ibañez et al., 2009, Pearson et al., 2004, Pearson et al., 2001, Wang et al., 2015a, Wang et al., 2015b). Pearson et al. (2001) used the spectral ratio 735/1005 nm for aflatoxin detection to optimally separate highly contaminated corn kernels (>100 ppb) from the less contaminated (<10 ppb). Based on reflectance spectra (500–1700 nm) of yellow maize kernels, Pearson et al. (2004) used 750 nm and 1200 nm to correctly identify over 99% of kernels as aflatoxin contaminated (>100 ppb) or uncontaminated. Tripathi & Mishra (2009) reported that bands related to fungal infection were found between 870 and 1200 nm and associated with Nsingle bondH bonds in most amino acids and aromatic rings. Wang et al., 2015a, Wang et al., 2015b indicated that wavelengths between 670.2 nm and 985.8 nm played a role in separating normal from pure AFB1 contaminated maize kernels, in addition, the authors identified 606.8, 671.6, 869.4, 917.6, 953.5 and 978.6 nm as key wavelengths for differentiating pure AFB1 contaminated maize kernels based on their concentrations of AFB1.

As mentioned, aflatoxin is a by-product of growth and metabolism of A. flavus in maize and other cereals. Correspondingly, to detect changes of maize nutrients and tissue structure caused by the growth of fungi can be taken as an indirect way for aflatoxin detection. Berardo et al. (2005) reported the possibility of quantifying the infection from fungi and metabolites produced in maize grain and flour by Fusarium verticilloides using NIR spectroscopy. Fernández-Ibañez et al. (2009) found that NIR spectroscopy was effective to detect aflatoxin presence at 20 ppb. A noteworthy observation is that there is limited literature related to the direct detection of fungal metabolites such as aflatoxin on the maize kernel surface by NIR spectroscopy. Hence, instead of detecting the metabolites produced by fungi in natural infected kernels, or the change in kernel nutritional composition of starch, proteins and lipids consumed by fungi, we would like to determine whether or not pure low level aflatoxin on/in clean maize kernels can be detected directly using VNIR hyperspectral imaging technology, and further we want to determine whether aflatoxin at concentrations as low as 10 ppb on/in clean maize kernels can be detected directly, and differentiated from the other aforementioned kernel attributes. Wang, Heitschmidt, et al. (2015) with VNIR reflectance measurement, showed that detection of pure aflatoxin artificially deposited on maize kernel surfaces was possible at concentrations as low as 10 ppb. However, Wang, Heitschmidt, et al. (2015) involved kernels from a single maize variety and the authors recommended that different maize varieties should be tested in order to build a more universal detection/classification model.

As stated above, most detection methods to date rely on indirect methods to detect aflatoxin, since the toxin's distribution within the kernel matrix does not allow for direct detection. Although in naturally contaminated maize kernels, the distribution of AFB1 is not uniform over the maize kernels and this does not allow for direct detection, previous studies such as Chu, Wang, Yoon, Ni, and Heitschmidt (2017) reported that AFB1 is mainly distributed on the germ part of the maize kernels. It is noteworthy that Chu et al. (2017), using SWIR hyperspectral imaging, detected AFB1 in maize kernels artificially inoculated with A. flavus in the field. The authors developed a distribution map to visualise the possible distribution of AFB1 in the maize samples and reported that even though the distribution of AFB1 was non-uniform in the kernels, the AFB1 was mainly distributed on the germ part of the kernels. From another study, Fernández-Ibañez et al. (2009) stated that AFB1 typically infects the maize kernel germ. The authors achieved natural infection of maize and barley kernels with growth of AFB1 at room temperature of 20 ± 2 °C for three months and analysed the samples with NIR spectroscopy. Pearson et al. (2001) reported that initially fungi such as A. flavus typically infects the kernel germ, uses the oil-rich germ for growth and metabolism and thus, the fungus should be more prevalent in the germ part. Therefore, it may be possible that the prevalence of fungi in the germ area may correlate to high distribution of aflatoxin in the germ part of the kernel, which was observed by Chu et al. (2017) and reported by Fernández-Ibañez et al. (2009).

The objective of our study was to examine the potential application of VNIR hyperspectral imaging to classify AFB1 contaminated maize kernels of four varieties from four representative major producing areas of the United States and get the spectra characteristics of AFB1, as well as to try to verify the possibility of direct detection of aflatoxin on maize kernels. Hence, by applying AFB1 solution to the germ-side surface of kernels and extracting average spectra, this present study, was envisaged to show how artificially AFB1 coated kernels can possibly contribute to a rapid aflatoxin detection method development. The use of average spectra extracted from the whole kernel germ-side region was expected to limit the influence of inhomogeneous distribution of AFB1 encountered in naturally infected kernels on the detection accuracy of AFB1. Considering the fact that conventional detection methods such as chromatography, and ELISA provide AFB1 concentration value measured basing on whole kernel, the average spectra of the corresponding kernel may have direct correlation to the AFB1 content in this kernel.

In this study, an experiment was conducted to demonstrate the feasibility of using VNIR hyperspectral imaging for directly detecting pure aflatoxin B1 coated onto maize kernel surfaces with the specific objectives of twofold: (1) to establish a classification model to classify AFB1 contaminated maize kernels of different varieties; and (2) to explain whether it is possible to directly detect the pure aflatoxin-coated maize kernel-surface using chemical information characterised by selected key wavelengths.

Section snippets

Maize kernels and sample preparation

Yellow Maize kernels of four varieties (Fig. 1) typically originating from different regions of the U.S.A., i.e., Georgia (GA), Illinois (IL), Indiana (IN) and Nebraska (NE) were investigated in this study. Details about the types of kernels and number of samples of each variety used in the study are presented in Table 1. Some differences between varieties include: (1) the IL and IN kernels are darker yellow than GA and NE; (2) the size of embryos are different and reduce in this order

Original spectra analysis

The raw mean kernel reflectance spectra for each variety are presented in Fig. 2a–d. The average absorption spectra of normal and AFB1 contaminated groups of each variety are illustrated in Fig. 3a–d. When compared to the normal kernels the average absorbance of contaminated kernels was higher for Georgia and Illinois varieties and lower for Indiana and Nebraska varieties in the visible region of around 400–600 nm. In addition, some differences in the spectra were revealed in the NIR range

Conclusions

The VNIR hyperspectral imaging technology and PCA/FDA statistical approach could differentiate uncontaminated and contaminated maize kernels with directly coated pure AFB1 and even discriminate between the contaminated kernels with different AFB1 levels. Detection of pure AFB1 artificially coated on maize kernel surfaces was possible at concentrations as low as 10 ppb. In addition, the study demonstrated the application of VNIR hyperspectral imaging system to examine the influence of varietal

Acknowledgement

The authors would like to thank Dr. Charles W Bacon, Research Leader of the Toxicology and Mycotoxin Research Unit, USDA, ARS for providing maize kernels. Ms. Peggy Feldner, Food Technologist, Mr. Vernon Savage, Engineering Technician, Mrs. Jerrie Barnett, Biological Laboratory Technician, and Ms. Candace Betts, Physical Science Technician with the Quality & Safety Assessment Research Unit, USDA, ARS, were also thanked for their assistance with sample preparation, fabricating the sample

References (39)

Cited by (67)

View all citing articles on Scopus
View full text