NON-DESTRUCTIVE IDENTIFICATION OF FUSARIUM INFECTED WHITE MAIZE KERNELS USING NIR HYPERSPECTRAL IMAGING AND MULTIVARIATE DATA ANALYSIS

White maize is a staple food in the diet of millions of people worldwide, including the majority of South Africans. To ensure the safety and quality of the crop, maize is graded according to South African legislation. In this legislation the maximum content of mouldy kernels is stipulated, where Fusarium spp. infections are the major cause. One great advantage to industry applications of near infrared (NIR) hyperspectral imaging, is overcoming the high risk of human error in manual classification processes. Human capacity to reproduce a consistent estimation of quality is limited, especially when several properties must be evaluated simultaneously, as in grading. It is the aim to automate the normally subjective grading process, also allowing it to become less tedious and time-consuming. Previous studies utilised sterilised kernels that had been inoculated post-harvest in a laboratory environment, where the consistency with field infections is disputed.
Multivariate classification models were developed using naturally infected maize kernels, encountered by graders at the South African Grain Laboratory (SAGL), Pretoria, South Africa. Images of Fusarium infected classes and sound classes were acquired using the SisuCHEMA short wave infrared camera in the wavelength range 900-2514 nm. Principal component analysis (PCA) with 6 PCs was calculated on mean-centred data in Evince v2.7.10 and irrelevant data were removed using brushing. In addition wavelengths from 900-1097 nm and 2477-2514 nm were removed. Savitsky-Golay filtering (3rd order polynomial, 2nd derivative, 15 points) and standard normal variate (SNV) transformation was applied to the data. The mean and difference spectra of the classes were studied in MATLAB v7.10. Using Evince, an object-wise PLS-DA model (66 objects) was calculated with full cross-validation.
In the cleaned PCA score plots of PC1 (79.6% SS) and PC2 (10.0% SS) distinct clusters for the kernel components (germ, floury endosperm and vitreous endosperm) were apparent, with no visual difference between the sound and infected classes. PC3 (2.6% SS) accounted for the difference between the up- and down-oriented germs of the kernels. This study aimed to find the more obscure differences associated with Fursarium infection. These were clouded by the more obvious differences between kernels, such as hardness variation, which many other studies have previously investigated. Thus, to reach beyond these sources of variation, the higher PCs were studied. In the score plot of PC1 (79.6% SS) vs PC5 (1.4% SS), there was a cluster that corresponded with the Fusarium infected class. This was evident in the score image of PC5.
Three prominent peaks were identified in the mean and difference spectra, namely 1198 and 1887 nm related to starch and 1430 nm related to protein. These constituents are associated with kernel components depleted during fungal proliferation, and are not associated with the fungal growth itself.
The object-wise PLS-DA model had satisfactory classification and prediction coefficients (R2 = 0.839; Q2 = 0.707), illustrating the potential for NIR hyperspectral imaging to be used as a rapid, objective method to assist graders in identifying Fusarium infected white maize. This could be broadened to other criteria in the South African grading regulation.