Non-destructive classification and prediction of aflatoxin-B1 concentration in maize kernels using Vis–NIR (400–1000 nm) hyperspectral imaging

Abstract

Aflatoxin-B1 contamination in maize is a major food safety issue across the world. Conventional detection technique of toxins requires highly skilled technicians and is time-consuming. Application of appropriate chemometrics along with hyperspectral imaging (HSI) can identify aflatoxin-B1 infected maize kernels. Present study was undertaken to classify 240 maize kernels inoculated with six different concentrations (25, 40, 70, 200, 300 and 500 ppb) of aflatoxin-B1 by using Vis–NIR HSI. The reflectance spectral data were pre-processed (multiplicative scatter correction (MSC), standard normal variate (SNV), Savitsky–Golay smoothing and their combinations) and classified using partial least square discriminant analysis (PLS-DA) and k-nearest neighbour (k-NN). PLS model was also developed to predict the concentration of aflatoxin-B1in naturally contaminated maize kernels inoculated with Aspergillus flavus. The potential wavelength (508 nm) was selected based on principal component analysis (PCA) loadings to distinguish between sterile and infected maize kernels. PCA score plots revealed a distinct separation of low contaminated samples (25, 40 and 70 ppb) from highly contaminated samples (200, 300 and 500 ppb) without any overlapping of data. The maximum classification accuracy of 94.7% was obtained using PLS-DA with SNV pre-processed data. Across all the combinations of pre-processing and classification models, the best efficiency (98.2%) was exhibited by k-NN model with raw data. The developed PLS model depicted good prediction accuracy ($R_{\mathit{CV}}^2$ = 0.820, SE_CV = 79.425, RPD_CV = 2.382) during Venetian-blinds cross-validation. The results of pixel-wise classification (k-NN) and concentration distribution maps (PLS with raw spectra) were quite close to the result obtained by reference method (HPLC analysis) of aflatoxin-B1 detection.

Introduction

Recent years have seen a steady rise in the cumulative annual growth rate of maize crop in India, at present it is 5.5 per cent (APEDA 2018). Growth in production or adoption of maize by farmers has both economic and legislative connotations. Increased production means increased storage needs. In a tropical country like India storage of food grains require massive planning and timely execution. Fungi, such as Aspergillus parasiticus and Aspergillus flavus infect stored maize and produce aflatoxins as their secondary metabolite (Shahin and Symons 2009). Also, there are evidences and reports to substantiate the prevalence of infection by fungi in maize kernels under biotic and abiotic stress before harvest in tropics and sub-tropics (Berardo et al. 2005). Aflatoxins are widely known for their carcinogenic potential; they are polycyclic unsaturated organic compounds and have a close relation with pentacyclic compounds. The carcinogenic properties are attributed to the presence of a lactone ring (Asao et al. 1963). Aflatoxins are reported to cause adverse health effects such as lung and liver cancer in humans and animal species, respectively (Shahin and Symons 2009). The grains with fungal growth acquire a different colour and a dusty texture; there is generation of a musty odour, discolouration, loss of germination and depletion of nutritional value and above all the capability of toxicity that the grain is poised to pose. Growth of fungi causes a rise in temperature resulting in a change of environment inside the storage structures, promoting the further transfer of this contamination to nearby grains. Aflatoxin contamination in grains results in massive economic loss to the farmers and raises serious food safety concerns for both humans and animals.

Food commodities have been known to contain aflatoxin, therefore countries all around the world have set a safe level of acceptance of aflatoxin B1 in the food substances, in India that level is 30 ppb (FSSR 2011) across all food items. Incidence of existence of aflatoxin B1 in maize was investigated through a multicenter survey by the Indian Council of Medical Research (ICMR) across eleven states representing the geographical variability of India. The survey found that out of the 2074 samples collected, 26% contained aflatoxin above the safe limit (Bhat et al. 1997). Chicken is the most favoured meat in India and its consumers are increasing at a rate of 10 per cent annually (Devi et al. 2014). Out of the total maize produced in India, about 47 per cent goes to the poultry feed industry. Another 20 and 14 per cent is consumed directly and as cattle feed, respectively. Early detection of fungal contamination shall prevent mycotoxin from entering the food chain through meat and milk. Aflatoxin detection techniques in vogue are gas chromatography (GC), high-performance liquid chromatography (HPLC) (Herzallah 2009) and thin layer chromatography (TLC) (Samarajeewa et al. 1991). These techniques require highly skilled technicians, sample preparation is costly and time consuming. The spectroscopic instruments are point-based scanning instruments; they lack information about the distribution or spread of the toxin over the substrate.

Spatial and spectral information of samples can be elicited by hyperspectral imaging (HSI), this would enable identification of infection at small affected areas as well as the spatial distribution of the infection caused by aflatoxin B1 on the maize kernels. HSI imaging technique obtains spectral data at each pixel and it correlates the same with neighbouring pixels also, thus, providing a large set (Vidal and Amigo 2012). Handling of such a huge amount of data and extraction of useful and significant information from the image is a tough task. Chemometrics is an attractive tool for the reduction in data dimensionality, extraction of essential information and qualitative or quantitative analysis of food (Ravikanth et al. 2017). Chemometric methods are mainly classified as exploratory analysis (PCA, ICA), supervised or unsupervised classification (ANN, PLS-DA, k-NN, etc.) and quantification (PCR, MLR, PLSR). The key features of chemometrics are its ability to reduce the complexity of huge hyperspectral data set, to develop prediction and classification models, and to enhance the accuracy and robustness of models based on spectral data analysis.

Hyperspectral imaging has been used in a plethora of incidences for food safety related issues of grains; Firrao et al. (2010) used NIR hyperspectral imaging in the range of 720–940 nm to identify fusarium infection in maize. Bauriegel et al. (2011) identified fusarium infection in wheat using Vis–NIR HSI (400–1000 nm). Ergot bodies in wheat kernels were identified with NRI HSI 900–1700 nm by Vermeulen et al. (2011). Other applications of HSI have been successful detection of undesirable substances in feed and food (Pierna et al. 2012), insect damaged wheat kernels (Singh et al. 2010), moisture content of barley, wheat and sorghum grain (McGoverin et al. 2011). Chemical and other optical properties of the grain are affected during contamination with aflatoxins which can be registered by visible or NIR spectroscopy (Wang et al. 2014).

Precise and rapid detection of aflatoxin in maize storage and handling is a challenge. Identification of wavelengths for the segregation of maize into infected and healthy is the first step in that direction. Thus, this study was conceived with the objective of identification of characteristic wavelength(s) to classify maize kernels on the basis of aflatoxin-B1 content by using vis–NIR hyperspectral imaging and development of a multivariate model that could predict the concentration of aflatoxin-B1 in maize kernels.

Materials and methods

Preparation of sample

In the present study 240 maize (NSC 1009) kernels were obtained from the Farm Section, ICAR-CIAE, Bhopal, India. The kernels had same pedigree and were harvested from the Institute farm in 2017. The maize kernels were treated by submerging in 0.8 per cent sodium hypochloride solution for 3 min to ensure that the kernels under experimentation did not have any pre-existing fungal contamination. Treated kernels were then washed 3–4 times with sterilized water to remove the residues of fungal growth and sodium hypochloride. The kernels were inoculated with six different levels of aflatoxin-B125, 40, 70, 200, 300, 500 ppb. These six concentrations of aflatoxin-B1 (RM4948, HiMedia Chemicals, Mumbai, India) stock solutions were prepared as based on the average mass of kernel (0.403 g, 12 per cent moisture). Requisite dilution was carried out by using methanol, the latter ensured elimination of residual mold spores, if any. The target concentration was achieved by the method elaborated by Wang et al. (2014). The kernels were placed in an ELISA microplate (24-well) as 6 columns and 4 rows, there was one kernel in each well. Every column had four kernels with the same concentration of toxin inoculated over it by a micropipette (Nichipet EX, Nichiryo, Japan). All the kernels were kept germ side down. Given the carcinogenic nature of aflatoxin B1, extreme caution was exercised while carrying out the inoculation and later during the imaging. Aspergillus flavus was inoculated to maize kernels and allowed to grow (Fig. 2d) under natural (30 °C, 75–80 per cent R_h) conditions for a period of 24 h, after which any further growth was restricted by keeping the kernels at 4 °C till it was further used. The HSI of these grains were used to validate the developed multivariate models. These grains were further used for measurement of aflatoxin-B1 destructively by high performance liquid chromatography (HPLC), these results were used to validate the performance of developed multivariate models.

Fig. 2.

Hyperspectral imaging spectral data analysis. a Spectral profile of maize kernels based on mean spectral data. b Principal component analysis for pixelwise classification (i) Loading plots for the three principal components (ii) Score surface plots PC₂ (iii) Image of control (dotted box) and aflatoxin infected germ down maize kernels (dashed line box) at 508 nm wavelength

HPLC conditions

HPLC (RF-10AXL, Shimadzu Corporation, Kyoto, Japan) system comprised an auto sampler, binary pump and column oven (CTO-10AS VP). Chromatography maize kernels was performed with a mobile phase (water: acetonitrile: methanol as 60:20:20, v/v/v) at a flow-rate of 0.5 mL/min. The analytical column (C18) was maintained at 40 °C. Fluorescence detection was performed at excitation and emission wavelengths of 350 and 450 nm, respectively. The chromatography method was partially adapted from Ghali et al. (2009).

Hyperspectral imaging system

The ELISA plates containing the twenty-four maize kernels were placed under Vis–NIR hyperspectral imaging system (OLES30, Specim, Oulu, Finland) comprising a digital camera (MV1-D1312, Photon focus AG, Switzerland), 25.4 mm diameter Long pass glass filter (Schott OG-590, Edmund Optics Inc., New Jersey, USA), spectrograph (IM Spector V10E, Specim, Spectral Imaging Ltd., Oulu, Finland) and three 50 W tungsten-halogen bulbs perched at an angle of 45^o with respect to the focus of camera lens. Setting up of the Specim spectral camera sCMOS parameters, real-time image visualization and data acquisition was carried out by Specim DAQ ver. 3.62 software. The hyperspectral imaging system had a spectral range of 398–1003 nm with a resolution of 6.237 nm and there were 97 bands for each pixel. This was a push-broom imaging system and acquired one line of the image at a time throughout the spectrum. In one line there were 1312 pixels, and the number of lines depended upon the length of the object.

Software

Reflectance correction was carried out by in-house generated scripts using the MATLAB (ver. 2018a) environment. The processing of the hyperspectral images and spectral data was performed by using HYPER-Tools v.2 (Amigo et al. 2015) which worked under MATLAB (The Mathworks, Inc., Natick, MA, USA) platform. Classification toolbox of MATLAB comprising supervised pattern recognition tools (Ballabio and Consonni 2013), Partial Least Square Discriminant Analysis (PLS-DA), and k-Nearest Neighbours (k-NN) was used for classification of samples based on the presence of varied concentration levels of aflatoxin-B1. HYPER-Tools v.2 (Amigo et al. 2015) was also used for development of Partial Least Square (PLS) models for prediction of aflatoxin-B1 concentration in naturally contaminated maize kernels inoculated with Aspergillus flavus.

Hyperspectral image processing

Calibration

The hyperspectral images for maize kernels were acquired in a reflectance mode. Reflectance calibration was performed to calibrate the raw intensity image into a reflectance image using black (about 0% reflectance) and white (about 99.9% reflectance) reference images. In order to remove the effect of dark current of the camera sensor, the black image (D_ref) was acquired when the light source was completely turned off and the camera lens was completely covered with its non-reflective opaque cap. The white reference image (W_ref) was obtained under the same condition as the raw image (R_image) using a white surface board which was a uniform, stable and high reflectance standard. The relative reflectance (R) was achieved for the hyperspectral image by:

$$R=\frac{R_{\mathit{image}}-D_{\mathit{ref}}}{W_{\mathit{ref}}-D_{\mathit{ref}}}\times 100$$ 1

Morphological operations and mean spectra extraction

The corrected hyperspectral image was spatially cropped to remove unnecessary parts not required for further analysis. The background of the image was removed by K-means clustering as the reflectance values of background and that of maize kernel varied widely. Besides background, masking also reduced the effects of non-uniform brightness and irregular shadows. Masking made the image a binary image with values 0 and 1, in the present research work the area with values 0 represented the background was ignored for the calculations, while values with area 1 represented the maize kernel. The average reflectance from each maize kernel was obtained by averaging all the pixels throughout the spectrum. Mean spectral data were used for the classification of maize kernels as per the aflatoxin-B1 content.

Data treatment

The obtained spectral data were pre-processed to counter the influence of certain undesirable phenomena arising during image acquisition, like-light scattering and spectral noise; and highlighting the differences between spectra for subsequent analysis. Image smoothing method was adopted for removal of noise from the instrument, while scattering was taken care of by multiplicative scatter correction (MSC) and standard normal variate (SNV). Combination of pre-processing operations was applied to mend the combined effects of noise and scattering (Ravikanth et al. 2017). Smoothing was carried out with a digital Savitsky-Golay (interval, 7 and order, 2) filter. MSC is quite frequently used as a pre-processing technique for scatter correction in NIR spectra. This technique reduces the scatter caused by the various constituents of the sample. Mathematically it can be represented by:

$$x_i^{\mathit{ORG}}=a_i+b_i{x}^{\mathit{REF}}+{\in }_i$$ 2

$$x_i^{\mathit{CORR}}=\frac{x_i^{\mathit{ORG}}-a_i}{b_i}=x^{\mathit{REF}}+uniquestructure$$ 3

where, x^ORG_i is original spectrum, x^REF_i is reference spectrum and x^CORR_i is corrected spectrum, a_i and b_i are correction coefficients of the ith sample. It can be evinced from Eqs. 2 and 3 that MSC does not eliminate scatter but decreases the inter-sample variance of the scatter by implementing an additive and multiplicative transformation of the individual spectrum into the simple idealized reference or mean spectrum. The SNV (Eq. 4), also known as a z-transformation, centering or scaling is another pre-processing technique for spectral data (Otto 1998). It aims to eliminate the multiplicative interferences caused as a result of sample particle size and its associated scatter (Buddenbaum and Steffens 2012):

$$x_i^{\mathit{CORR}}=\frac{x_i^{\mathit{ORG}}-a_i}{b_i}$$ 4

where x^CORR_i is SNV corrected and x^ORG_i is original spectrum, a_i and b_i are mean and standard deviation of the ith sample, respectively. The methodology has been depicted in Fig. 1 for better understanding.

Fig. 1.

Comprehensive flow-chart of the methodology for hyperspectral image analysis of aflatoxin-B1 inoculated maize kernels

Principal component analysis (PCA)

PCA is a well known method of multivariate analysis (Kemsley 1996) and is used to transform a large dataset into smaller dimension dataset in which there are a number of inter-correlated variables while retaining most of the information that the dataset aims to convey. PCA is generally used for data dimension reduction, elimination of multi collinearity and enhancement of the hyperspectral dataset to ease feature selection (Romero 2010). Identificationof the wavelength representing the presence of aflatoxin-B1 in the maize kernel was made with the help of PCA.

A typical hypercube of a hyperspectral dataset (H) has the dimensions as X × Y × λ, and can be decomposed using a PCA into a set of scores and loading (Amigo 2010). The matrix of the transformed dataset can be represented as:

$$H^I=S{L}^S+E$$ 5

where H^I is the transformed dataset having a dimension XY × λ, S is the score surface with dimension XY × F, L^S is the loading profiles with dimension F × λ and E is the residual matrix with dimension XY × λ.

Supervised classification

Supervised learning has widespread successful application in hyperspectral imaging of food and agricultural applications (Ravikanth et al. 2017). In this research work, all the developed models were cross-validated by using Venetian blinds approach with 5 groups.

Partial least square-discriminant analysis (PLS-DA)

It combines the qualities of PLS and DA which is regression power and classification power, respectively. PLS-DA has been successfully used for hyperspectral imaging data analysis (Vermeulen et al. 2011). In PLS-DA every correctly classified item will be given a 1 and wrongly classified items will be identified as 0. The classification of maize kernels on the basis of concentration of aflatoxin B1 was carried out by this technique. The latent variables for each data treatment technique were selected based on the minimum error rate and the number of not assigned samples during cross-validation.

k-Nearest neighbour (k-NN)

k-NN is a supervised classification algorithm. Here, k denotes the number of closest objects that shall be classified with respect to k-nearest neighbours in the data space. It is useful for multiclass problems and it is a non-linear and parametric method (Ballabio and Todeschini 2009). The performance of the model is influenced by the k-value (Kong et al. 2013). Classification accuracy of the model depends on the dataset (Zheng et al. 2014). The k-NN model was developed to classify the maize kernels based on the aflatoxin-B1 concentration. The k-value for each data treatment technique was selected based on the minimum error rate.

Partial least square (PLS) regression model

PLS modeling finds excellent application where the number of independent variables (here, wavelengths) is much higher than the number of dependent variables (Xiaobo et al. 2010). PLS captures the variance and correlates it with the data, all this while the covariance is maximized. Before the onset of model development PLS decomposes the X and Y data into scores and loadings. The X data matrix (spectral reflectance) is decomposed into a matrix X_P (scores) and a matrix P (loadings) plus an error (E₁). The matrix Y (aflatoxin-B1 content) is decomposed into Y_P (scores) and Q (loadings) and the error (E₂). Using all these factors a PLS model is developed between scores of X and Y, i.e. X_P and Y_P respectively.

$$X=X_p{P}^T+E_1$$ 6

$$Y=Y_p{Q}^T+E_2$$ 7

The goal of the PLS algorithm is to minimize the norm of E₂ while maintaining the correlation between X and Y by the inner relation,

$$Y_P=W{X}_P$$ 8

The aflatoxin content of maize kernels was estimated by the NIPALS algorithm with the whole raw spectral range. The developed model was internal cross-validated with Venetian blind cross-validation with five groups to avoid over-fitting of the model. The performance of developed model was tested with maize kernels inoculated with Aspergillus flavus and allowed to grow under natural (30 °C, 75–80% R_h) conditions for a period of 24 h.

Statistical assessment of data

Classification model

The statistical validation of classification model performance was carried out by using classifiers, sensitivity, specificity and class error.

Sensitivity was defined as the ability of a model to correctly classify the samples belonging to the particular class.

$$Sensitivity=\frac{\mathit{TP}}{TP+FN}$$ 9

Specificity was the capability of a model to reject the samples of other classes.

$$Specificity=\frac{\mathit{TN}}{FP+TN}$$ 10

$$Errorrate=1-\left(\frac{Sensitivity+Specificity}{2}\right)$$ 11

$$Nonerrorrate=\frac{Sensitivity+Specificity}{2}$$ 12

where, TP is true positive enumerating the samples that are correctly classified as belonging and TN is true negative enumerating the samples that are correctly assigned as not belonging, to a particular class. FP is false positive, samples that are incorrectly classified as belonging and FN is false negative, samples assigned incorrectly as not belonging, to a particular class (Ballabio and Consonni 2013; Amigo et al. 2015). Accuracy is the proportion of correctly assigned samples to the total number of samples (excluding not assigned samples) (Ballabio and Consonni 2013).

Prediction model

The statistical performance of the developed PLS model was evaluated by - coefficient of determination of cross-validation (R²_CV), and standard error of cross-validation (SE_CV), bias and residual predictive deviation of cross-validation (RPD_CV). In order to standardize the SE_CV, other statistical parameter RPD_CV was calculated as the ratio of standard deviation (SD) of calibration data set to the SE_CV (Sánchez et al. 2018). Gaston et al. (2010) reported that RPD value > 2.5 means the model performance was excellent, if RPD lies between 2.0 and 2.5 the model can be designated as very good, an RPD value between 1.8 and 2.0 indicates a good model, a model is fair if the RPD lies between 1.4 and 1.8, poor models have RPD between 1 and 1.4, a model cannot be applied if RPD is less than 1.

Results and discussion

Hyperspectral images of healthy and infected maize kernels were corrected using the ‘dark’ and ‘white’ references. Background removal and other morphological operations were performed so that the subtle changes in the image due to the presence of varied concentration of aflatoxin B1 could be registered.

Spectral profile

The reflectance spectra encompassing the spectral region of 398–1003 nm for all the seven experimental groups, control, 25, 40, 70, 200, 300 and 500 ppb is presented in Fig. 2a. It can be seen that spectra was a bit jagged at the extreme ends of the spectrum, this was due to low signal to noise ratio of the system, which had been stable for all the observations. The morphological features of the curves changed markedly for those that represented the six concentrations of aflatoxin B1 contamination. It can be inferred by visual inspection of mean spectra based on the level of aflatoxin B1 concentration in maize Fig. 2a that the spectral lines demonstrate a steep ascent at line ‘A’ (approximately, 510 nm) and this trend continued till they plateau at line ‘B’ (approximately, 600 nm). Between these two lines deviation kept on varying between the spectral line representing a specific concentration of aflatoxin B1 inoculation on the maize kernel. Further it can be inferred that the reflectance of contaminated samples decreased with increase in the aflatoxin B1 concentration.

Dimensionality reduction based on masking and PCA

The first three principal components explained a total variability of 93.84 per cent (PC₁ = 84.47 per cent, PC₂ = 5.91 per cent, PC₃ = 3.46 per cent), graphical representation is exhibited as Fig. 2bi. While PC₁ indicated mainly about the endosperm and other contents of kernels, PC₃ showed the appearance or texture characteristics of the kernels. The second principal component (PC₂) featured a steep rise and fall at wavelengths identified as 508 nm and 580 nm. These wavelengths matched to the location of mutations in the original spectra (Fig. 2a). Also, the surface score plot (Fig. 2bii) represents the remarkable segregation between healthy and infected kernels. Capturing the snapshot of a hyperspectral image at 508 nm revealed a distinct classification of kernels as healthy and infected (Fig. 2biii). However, at this wavelength segregation of grains on the basis of individual aflatoxin-B1 content could not take place; similar observations were made by Wang et al. 2014.

Exploratory analysis of kernels (PCA)

Generally, only a few PC’s are enough to capture the variability of the whole data set while not losing out on any important information. The scree plot (Fig. 3a) shows the eigenvalues in the y-axis and the number of principal components on the x-axis. The inspection of the scree plot indicated that the downward curve structures an elbow right after the third principal component, marked by a dashed-line rectangular box.

Fig. 3.

Principal component analysis for individual kernels. a Scree plot depicting the selection of the number of components. b Score plot representing clustering of the maize kernel samples based on aflatoxin-B1 concentration for 240 samples

The first three principal components explained a total variability of 98.37% (PC₁ = 84.46 per cent, PC₂ = 11.12 per cent, PC₃ = 2.78 per cent). The score plot of PC₁ and PC₂ is illustrated in Fig. 3b. It can be evinced that PC₁ was able to represent a clear distinction between the grains vis-a-vis their aflatoxin-B1 concentration. The lower aflatoxin B1 concentration (25, 40 and 70 ppb) kernels had negative scores and remaining kernels (200, 300 and 500 ppb) had positive scores. The score values of PC₁ increased with an increase in aflatoxin B1 concentration. Concentrations of 25 and 40 ppb had the almost same scores (both the samples overlapped each other) and it was difficult to classify those samples according to their aflatoxin B1 concentration. Misclassification for lower aflatoxin B1 concentration was also reported by Kandpal et al. (2015) with maize kernels. However, 70 ppb was markedly separated from 25 and 40 ppb. Higher concentration kernels (200, 300 and 500 ppb) were classified without any overlapping among each other. Some of the 500 ppb kernels could be spotted with 200 and 300 ppb kernels, it may be due to the uneven spatial distribution of aflatoxin B1 over the kernel surface during the artificial inoculation, or the selected pixels contained an undefined concentration of aflatoxin B1. Similar results were reported by Pearson et al. (2001) using spectral reflectance ratio 735/1005 nm, Del Fiore et al. 2010 classified different levels of toxigenic fungi in maize at 870 nm and Wang et al. 2015 used Vis/NIR hyperspectral imaging and a PCA/FDA statistical approach to classify artificially inoculated maize kernels for concentrations as low as 10 ppb.

Supervised classification

The optimum number of latent variables of PLS-DA model for different pre-processing techniques was selected based on the criteria of minimum error rate and minimum non- assigned samples. Similarly, k-value of k-NN classification for different pre-processing methods was selected based on the minimum error rate. Raw and different pre-processing techniques functioned with optimum latent variables for PLS-DA model, k-value for k-NN model (Raw: 14, 2; MSC: 16, 3; SNV: 12, 3; MSC + SNV: 12, 3; MSC + Smoothing: 18, 2; SNV + Smoothing: 17, 2). The performance of classification techniques as gauged by different classifiers is reported in Table 1. The ‘sensitivity’ and ‘specificity’ of 25, 70, 300 and 500 ppb kernels was higher than 0.95 for all pre-processing techniques, except for 40 and 200 ppb during PLS-DA classification.

Table 1.

Performance of classification methods during cross validation and training for various data pre-processing techniques by different classifiers

Pre-processing	Classifiers	PLS-DA						k-NN
		25 ppb	40 ppb	70 ppb	200 ppb	300 ppb	500 ppb	25 ppb	40 ppb	70 ppb	200 ppb	300 ppb	500 ppb
Raw	SE	1 (1)	0.73 (0.91)	0.93 (0.94)	0.83 (0.94)	0.97 (1)	0.89 (0.97)	0.97 (1)	1 (1)	1 (1)	0.97 (0.97)	0.97 (0.97)	0.9 (0.9)
	SP	0.98 (0.99)	0.98 (0.98)	0.97 (0.99)	0.97 (1)	0.97 (0.99)	1 (1)	1 (1)	1 (1)	1 (1)	0.99 (0.99)	0.99 (0.99)	0.99 (0.99)
	NER	0.99 (0.995)	0.86 (0.95)	0.95 (0.96)	0.9 (0.97)	0.97 (0.99)	0.95 (0.98)	0.98 (1)	1 (1)	1 (1)	0.98 (0.98)	0.98 (0.98)	0.95 (0.95)
	ER	0.01 (0.005)	0.14 (0.05)	0.05 (0.04)	0.1 (0.03)	0.03 (0.005)	0.05 (0.02)	0.02 (0)	0 (0)	0 (0)	0.02 (0.02)	0.02 (0.02)	0.055 (0.05)
MSC	SE	0.96 (1)	0.88 (0.96)	0.94 (0.97)	0.74 (0.89)	0.97 (0.97)	0.96 (1)	0.82 (0.88)	0.68 (0.65)	0.93 (0.95)	0.78 (0.78)	0.72 (0.75)	0.68 (0.65)
	SP	0.99 (1)	0.97 (0.98)	0.96 (0.99)	0.98 (1)	0.99 (0.99)	1 (1)	0.96 (0.96)	0.94 (0.95)	0.91 (0.91)	0.94 (0.95)	0.98 (0.98)	0.98 (0.97)
	NER	0.98 (1)	0.93 (0.97)	0.95 (0.98)	0.86 (0.95)	0.98 (0.98)	0.98 (1)	0.89 (0.92)	0.81 (0.8)	0.92 (0.93)	0.86 (0.86)	0.85 (0.86)	0.83 (0.81)
	ER	0.02 (0)	0.07 (0.03)	0.05 (0.02)	0.14 (0.05)	0.02 (0.02)	0.02 (0)	0.11 (0.08)	0.19 (0.2)	0.08 (0.07)	0.14 (0.14)	0.15 (0.13)	0.17 (0.19)
SNV	SE	1 (1)	0.89 (0.92)	0.93 (1)	0.73 (0.92)	0.92 (1)	1 (0.96)	0.82 (0.88)	0.68 (0.65)	0.93 (0.95)	0.78 (0.78)	0.72 (0.75)	0.68 (0.65)
	SP	0.99 (1)	0.98 (0.99)	0.94 (0.98)	1 (1)	0.96 (0.98)	1 (1)	0.96 (0.96)	0.94 (0.95)	0.91 (0.91)	0.94 (0.95)	0.98 (0.98)	0.98 (0.97)
	NER	0.995 (1)	0.94 (0.95)	0.94 (0.99)	0.87 (0.96)	0.94 (0.99)	1 (0.98)	0.89 (0.92)	0.81 (0.8)	0.92 (0.93)	0.86 (0.86)	0.85 (0.86)	0.83 (0.81)
	ER	0.005 (0)	0.06 (0.05)	0.06 (0.01)	0.13 (0.04)	0.06 (0.01)	0 (0.02)	0.11 (0.08)	0.19 (0.2)	0.08 (0.07)	0.14 (0.14)	0.15 (0.14)	0.17 (0.19)
MSC + SNV	SE	1 (1)	0.89 (0.92)	0.93 (1)	0.73 (0.92)	0.92 (1)	1 (0.96)	0.82 (0.88)	0.68 (0.65)	0.93 (0.95)	0.78 (0.78)	0.72 (0.75)	0.68 (0.65)
	SP	0.99 (1)	0.98 (0.99)	0.94 (0.98)	1 (1)	0.96 (0.98)	1 (1)	0.96 (0.96)	0.94 (0.95)	0.91 (0.91)	0.94 (0.95)	0.98 (0.98)	0.98 (0.97)
	NER	0.995 (1)	0.94 (0.95)	0.94 (0.99)	0.84 (0.96)	0.94 (0.99)	1 (0.98)	0.89 (0.92)	0.81 (0.8)	0.92 (0.93)	0.86 (0.86)	0.85 (0.86)	0.83 (0.81)
	ER	0.005 (0)	0.06 (0.05)	0.06 (0.01)	0.14 (0.04)	0.06 (0.01)	0 (0.02)	0.11 (0.08)	0.19 (0.2)	0.08 (0.07)	0.14 (0.14)	0.15 (0.14)	0.17 (0.19)
MSC + S	SE	0.92 (0.96)	0.78 (0.88)	1 (0.97)	0.75 (0.85)	0.97 (1)	1 (1)	0.8 (0.82)	0.68 (0.7)	0.95 (0.95)	0.8 (0.8)	0.85 (0.88)	0.78 (0.75)
	SP	0.98 (0.99)	0.98 (0.99)	0.94 (0.97)	0.98 (0.99)	1 (1)	1 (1)	0.96 (0.96)	0.94 (0.94)	0.96 (0.96)	0.95 (0.96)	0.98 (0.98)	0.97 (0.97)
	NER	0.95 (0.975)	0.88 (0.93)	0.97 (0.97)	0.86 (0.92)	0.99 (1)	1 (1)	0.88 (0.89)	0.81 (0.82)	0.95 (0.95)	0.88 (0.88)	0.92 (0.93)	0.88 (0.86)
	ER	0.05 (0.025)	0.12 (0.07)	0.03 (0.03)	0.14 (0.08)	0.01 (0)	0 (0)	0.12 (0.11)	0.19 (0.18)	0.05 (0.05)	0.12 (0.12)	0.08 (0.07)	0.12 (0.14)
SNV + S	SE	0.92 (0.96)	0.78 (0.96)	1 (0.97)	0.73 (0.85)	1 (1)	1 (1)	0.8 (0.82)	0.7 (0.7)	0.95 (0.95)	0.8 (0.8)	0.85 (0.88)	0.78 (0.75)
	SP	0.98 (0.99)	0.97 (0.99)	0.95 (0.99)	0.99 (1)	1 (0.99)	1 (1)	0.96 (0.96)	0.94 (0.94)	0.96 (0.96)	0.95 (0.96)	0.98 (0.98)	0.97 (0.97)
	NER	0.95 (0.97)	0.88 (0.98)	0.98 (0.98)	0.86 (0.93)	1 (0.995)	1 (1)	0.88 (0.89)	0.82 (0.82)	0.95 (0.95)	0.88 (0.88)	0.92 (0.93)	0.88 (0.86)
	ER	0.05 (0.03)	0.12 (0.02)	0.02 (0.02)	0.14 (0.07)	0 (0.005)	0 (0)	0.12 (0.11)	0.18 (0.18)	0.05 (0.05)	0.12 (0.12)	0.08 (0.07)	0.12 (0.14)

MSC Multiplicative scatter correction, SNV Standard normal variate, S-SG smoothing, PLS-DA Partial least square discriminate analysis, k-NN k Nearest neighbour, SE Sensitivity, SP Specificity, NER non error rate and ER-Error rate

Values in the parenthesis are the classifier values for the training dataset

The lower sensitivity and specificity of 40 and 200 ppb samples was due to the misclassification of samples. The performance indicators for cross- validation as well as training for specific classifiers under respective pre-processing techniques were distinctly similar in value, indicating the fact that the developed models were free from over and underfitting. The ‘specificity’ and ‘sensitivity’ of k-NN model for raw data were higher than 0.95 for all the classifiers for raw data. This means that there was a chance of 95 per cent of kernels be classified correctly to the corresponding category. The non error rate for all the classes was higher than 0.95.

The maximum classification accuracy of 94.7 per cent was obtained using PLS-DA with SNV pre-processing. Across all the combinations of pre-processing techniques and classification models, the efficiency of k-NN model while using raw data was maximum at 98.2 per cent. Again, performance of PLS-DA model was limited by a sizeable number of non-assigned (NA) entities, whereas there was no such issue with k-NN model (Table 2). Even if PLS-DA model needs to be considered then the data should be pre-processed with SNV, in that case also NA will be high for low aflatoxin B1 concentration levels (Table 2). Based on the statistical indices it can be concluded that k-NN performed far better than PLS-DA. The k-NN model developed with raw spectral performed better than the other pre-processing techniques. Therefore, the k-NN model developed with raw spectral data was used for pixel-wise classification of the maize kernels where Aspergillus flavus was grwon naturally (Fig. 4a), the pixels were collated based on the experimental range (25, 40, 70, 200, 300, and 500 ppb) of the aflatoxin-B1 concentration. Based on the spatial distribution of the contamination as identified by k-NN model categorised into four concentration levels, it was predicted that aflatoxin-B1 content is 178 ppb (Fig. 4a). This is slight deviation from the result of the destructive analysis, which is 163 ppb (Fig. 4b). The algorithm of k-NN works by clubbing the similar pixels (here, of same aflatoxin-B1 concentration) together, while doing so some borderline cases are misappropriated resulting in an error ridden statistically relevant prediction.

Table 2.

Confusion matrix dipicting performance of different classification methods against various data pre-processing techniques

Pre -processing	Aflatoxin concentration	PLS-DA							k-NN
		Predicted							Predicted
		25 ppb	40 ppb	70 ppb	200 ppb	300 ppb	500 ppb	NA	25 ppb	40 ppb	70 ppb	200 ppb	300 ppb	500 ppb	NA
Raw
Actual	25 ppb	23 (30)						17 (10)	39 (40)	1 (0)					0 (0)
	40 ppb	3 (1)	24 (30)	4 (2)	2 (0)			7 (7)		40 (40)					0 (0)
	70 ppb		1 (1)	26 (31)		1 (1)		12 (7)			40 (40)				0 (0)
	200 ppb		2 (2)	1 (0)	24 (34)	2 (0)		11 (4)				39 (39)		1 (1)	0 (0)
	300 ppb				1 (0)	32 (33)		7 (7)					39 (39)	1 (1)	0 (0)
	500 ppb				2 (0)	1 (1)	25 (30)	12 (9)				2 (2)	2 (2)	36 (36)	0 (0)
MSC
Actual	25 ppb	25 (32)		1 (0)				14 (8)	33 (35)	4 (2)		1 (1)	1 (1)	1 (1)	0 (0)
	40 ppb	1 (0)	21 (24)	1 (1)	1 (0)			16 (15)	4 (5)	27 (26)	4 (5)	3 (1)		2 (3)	0 (0)
	70 ppb			29 (34)	1 (0)	1 (1)		9 (5)			37 (38)	2 (1)	1 (1)		0 (0)
	200 ppb	1 (0)	4 (3)	2 (0)	20 (25)			13 (12)	1 (1)	2 (2)	4 (4)	31 (31)	2 (2)		0 (0)
	300 ppb			1 (1)		29 (34)		10 (5)	1 (1)	1 (1)	6 (6)	2 (1)	29 (30)	1 (1)	0 (0)
	500 ppb				1 (0)		24 (30)	15 (10)	1 (1)	5 (4)	4 (4)	3 (5)		27 (26)	0 (0)
SNV
Actual	25 ppb	17 (23)						23 (17)	33 (35)	4 (2)		1 (1)	1 (1)	1 (1)	0 (0)
	40 ppb	1 (0)	17 (24)	1 (2)				21 (14)	4 (5)	27 (26)	4 (5)	3 (1)		2 (3)	0 (0)
	70 ppb			28 (32)		2 (0)		10 (8)			37 (38)	2 (1)	1 (1)		0 (0)
	200 ppb		2 (1)	3 (0)	19 (23)	2 (1)		14 (15)	1 (1)	2 (2)	4 (4)	31 (31)	2 (2)		0 (0)
	300 ppb			2 (0)		24 (28)		14 (12)	1 (1)	1 (1)	6 (6)	2 (1)	29 (30)	1 (1)	0 (0)
	500 ppb					0 (1)	19 (25)	21 (14)	1 (1)	5 (4)	4 (4)	3 (5)		27 (26)	0 (0)
MSC + SNV
Actual	25 ppb	17 (23)						23 (17)	33 (35)	4 (2)		1 (1)	1 (1)	1 (1)	0 (0)
	40 ppb	1 (0)	17 (24)	1 (2)				21 (14)	4 (5)	27 (26)	4 (5)	3 (1)		2 (3)	0 (0)
	70 ppb			28 (32)		2 (0)		10 (8)			37 (38)	2 (1)	1 (1)		0 (0)
	200 ppb		2 (1)	3 (0)	19 (23)	2 (1)		14 (15)	1 (1)	2 (2)	4 (4)	31 (31)	2 (2)		0 (0)
	300 ppb			2 (0)		24 (28)		14 (12)	1 (1)	1 (1)	6 (6)	2 (1)	29 (30)	1 (1)	0 (0)
	500 ppb					0 (1)	19 (25)	21 (14)	1 (1)	5 (4)	4 (4)	3 (5)		27 (26)	0 (0)
MSC + S
Actual	25 ppb	24 (26)		2 (1)				14 (13)	32 (33)	4 (3)		2 (2)	1 (1)	1 (1)	0 (0)
	40 ppb	1 (0)	21 (23)	3 (2)	2 (1)			13 (14)	6 (5)	27 (28)	4 (4)	2 (1)	0 (1)	1 (1)	0 (0)
	70 ppb		0 (1)	26 (29)				14 (10)		2 (2)	38 (38)				0 (0)
	200 ppb	2 (2)	3 (1)	2 (1)	21 (23)			12 (13)	1 (1)	3 (3)		32 (32)	2 (2)	2 (2)	0 (0)
	300 ppb			1 (0)		28 (31)		11 (9)	1 (1)	1 (1)	1 (1)	2 (1)	34 (35)	1 (1)	0 (0)
	500 ppb						21 (28)	19 (12)		3 (3)	2 (3)	4 (4)		31 (30)	0 (0)
SNV + S
Actual	25 ppb	22 (25)		2 (1)				16 (14)	32 (33)	4 (3)		2 (2)	1 (1)	1 (1)	0 (0)
	40 ppb	2 (0)	21 (22)	3 (1)	1 (0)			13 (17)	6 (5)	28 (28)	4 (4)	1 (1)	0 (1)	1 (1)	0 (0)
	70 ppb		0 (1)	24 (30)				16 (9)		2 (2)	38 (38)				0 (0)
	200 ppb	1 (2)	4 (1)	2 (0)	19 (23)	0 (1)		14 (13)	1 (1)	3 (3)		32 (32)	2 (2)	2 (2)	0 (0)
	300 ppb					30 (33)		10 (7)	1 (1)	1 (1)	1 (1)	2 (1)	34 (35)	1 (1)	0 (0)
	500 ppb						22 (28)	18 (12)		3 (3)	2 (3)	4 (4)		31 (30)	0 (0)

Fig. 4.

Performance of k-NN and PLS model. a Pixel-wise classification of maize kernels using k-NN model. b Result of reference method (HPLC analysis) for aflatoxin-B1 content in maize kernels. c Actual and predicted aflatoxin–B1 content using PLS model during calibration and cross validation. d Concentration distribution map developed using PLS model

PLS model validation

The PLS regression model was developed using full raw spectral data. The optimum number of LVs was selected as 10 from the plot between LVs and RMSE during Ventain blind cross validation (5 groups) technique. The plot between actual and predicted aflatoxin-B1 content of maize kernels during model calibration and cross-validation is shown in Fig. 4c. The statistical results obtained during model calibration (R²_C = 0.873; SE_C = 65.082; slope_C = 0.873) and cross validation (R²_CV = 0.820; SE_CV = 79.425; slope_CV = 0.820) depicted that the model was free from over and underfitting. Further, the efficacy of the model was established by RPD_CV (2.382) value (Gaston et al., 2010). The slope of regression line during model cross validation (red line) was almost close to the slope of the actual regression line (black line). Some of the data points whose concentration was below the standard error of the model were assigned to negative values (Fig. 4c). The developed model was applied to HSI of maize kernels where Aspergillus flavus was grown under natural conditions, the contamination distribution map of each and every grain is shown in Fig. 4d. It can be elicited that most of the pixels fall within the aflatoxin-B1 concentration range of 175-185 ppb as predicted by statistically valid PLS model.

Conclusion

Vis–NIR hyperspectral imaging can be used for the segregation of healthy samples from the contaminated samples. The featured wavelength for segregating the fresh samples from contaminated samples is 508 nm. The exploratory analysis (PCA) results revealed a clear separation of low contaminated samples (25, 40 and 70 ppb) from highly contaminated samples (200, 300 and 500 ppb) was possible without any overlapping of data, but there was an overlapping of points within the ranges of low contamination and high contamination. The spectral data without any pre-processing with k-NN classifier yielded maximum classification rate at 98.3 per cent, this was the best combination of pre-processing and classification technique to classify the maize kernels with respect to the extent of aflatoxin B1 contamination. The developed PLS model with raw spectral data exhibited very good prediction accuracy and it can be used effectively to predict the aflatoxin-B1 content in maize kernels within the experimental range (25–500 ppb) rapidly with low error. Vis–NIR hyperspectral imaging with aforesaid chemometrics shall be successful in the classification of maize kernels on the basis of aflatoxin-B1 contamination and also prediction of the extent of contamination.