Reduced wavelength spectral imaging for grading defect and asymptomatic Fusarium detection in white maize

Sendin, Kate (2020-04)

Thesis (PhD)--Stellenbosch University, 2020.

Thesis

ENGLISH ABSTRACT: The aim of this dissertation was to present the South African maize industry with an accurate and affordable automated analytical technique for white maize grading and for identifying asymptomatic Fusarium fungal contamination. This was achieved by using near infrared (NIR) hyperspectral imaging, chemometric classification model development, optimal waveband selection and hierarchical modelling. White maize grade is assigned based on the content of 5 main categories in a maize consignment, namely sound white maize, defective white maize, pinked white maize, yellow maize and foreign materials. Defective white maize and foreign materials comprise further sub-categories, giving a total of 17 classes. All of the categories stipulated in South African maize grading legislation were simultaneously classified (1044 samples; 60 kernels of each class) using NIR hyperspectral imaging and partial least squares discriminant analysis (PLS-DA) models assembled in a hierarchical decision pathway. The hierarchical model divided the task into 25 small steps (binary and ternary PLS-DA models), which progressed from the most easily classified classes to the most difficult. The hierarchical model was based on the full NIR spectrum (288 wavebands) and performed with an overall accuracy of 93.3% for the main categories. The classification of sound white maize (88.3%), pinked white maize (83.3%) and yellow maize (75.0%) should ideally be improved before the method is implemented for industry grading. Pinked white maize and yellow maize are distinguishable due to the presence of anthocyanin and beta-carotene, respectively, which both exhibit maximum absorption in the visible region and do not interact with NIR radiation. The use of a spectral imaging system including the visible region is expected to improve the classification of these classes. Following the encouraging success of maize grading using the full NIR spectrum, waveband reduction and optimisation was conducted to attempt simplified but accurate grading of white maize using a recalculated hierarchical decision pathway. Three waveband selection techniques were employed, namely waveband windows (48 wavebands), variable importance in projection (VIP) (21 wavebands) and covariance selection (CovSel) (13 wavebands). There was a loss of performance in all three reduced waveband models. The waveband windows (87.1% main category classification accuracy) and VIP waveband sets (84.5% main category classification accuracy) performed with similar classification accuracies across the numerous categories, but the VIP waveband set utilised less than half of the spectral variables. The CovSel waveband set used the fewest wavebands but exhibited an unacceptable loss of classification accuracy (81.9% main category classification accuracy). Overall, the VIP waveband set (964, 1127, 1159, 1323, 1356, 1388, 1421, 1716, 1847, 1879, 1912, 1945, 2043, 2239, 2272, 2305, 2337, 2403, 2435, 2468 and 2501 nm), which was based on only 7.3% of the 288 original spectral variables, was recommended as the best trade-off between instrument performance and expected cost of the system. A second issue in the South African maize industry was addressed, namely the detection of single asymptomatically Fusarium infected kernels. NIR hyperspectral images of 224 visibly sound (healthy) kernels were acquired prior to germination of the kernels in individual sterile containers. Germination caused internal Fusarium infections to become visibly identifiable as external fungal growth, which was later confirmed by conventional microbial testing. While only 3.3% of the kernels in the bulk samples exhibited visible rotting symptoms (flagged during visual inspection), 32.1% of germinated kernels were asymptomatically infected and capable of producing harmful fumonisin mycotoxins. Some of these bulk samples contained fumonisin levels of 8 ppm (double the limit) but would have been declared safe for human consumption based on manual inspection methods. This lack of correlation between visible symptoms and safety emphasised the need for additional analytical methods to determine Fusarium related risks. The pre-germination spectral images of the uninfected and asymptomatically infected kernels were divided into two classes, and a PLS-DA model classified the maize kernels with a classification accuracy of 67.0%. Considering the high food safety risk associated with fumonisins, NIR hyperspectral imaging is not a viable method for detecting asymptomatic Fusarium infections during South African white maize processing. The results of this study demonstrated that NIR hyper- and multispectral imaging are promising analytical techniques for automated maize grading, but not for the detection of asymptomatic Fusarium infection. However, the results of the Fusarium germination study provided insight into the status of Fusarium infection and fumonisins in the South African maize industry that have not yet appeared in literature and emphasised the need for industry-friendly mycotoxin testing methods.

AFRIKAANSE OPSOMMING: Die doel van hierdie proefskrif was om die Suid-Afrikaanse mieliebedryf voor te stel aan 'n akkurate en bekostigbare outomatiese analitiese tegniek vir witmielie-gradering en die identifisering van asimptomatiese Fusarium-swambesmetting. Dit is bewerkstellig deur gebruik te maak van naby-infrarooi (NIR) hiperspektrale beelding, chemometriese klassifikasies modelontwikkeling, optimale golfband selektering en hiërargiese modellering. Witmielie-graad word op grond van die inhoud van vyf hoofkategorieë in 'n mieliebesending, naamlik gesonde witmielies, foutiewe witmielies, pienk witmielies, geelmielies en vreemde materiale toegeken. Foutiewe witmielies en vreemde materiale bestaan uit verdere subkategorieë, wat 'n totaal van 17 klasse gee. Al die kategorieë, soos uiteengesit in Suid-Afrikaanse mielie-graderingswetgewing, is gelyktydig geklassifiseer (1044 monsters; 60 pitte van elke klas) deur gebruik te maak van NIR-hiperspektrale beeldvorming en parsiële kleinste kwadrate-diskriminantanalise (PLS-DA) modelle wat in 'n hiërargiese besluitweg saamgestel is. Die hiërargiese model het die taak in 25 klein stappe (binêre en ternêre PLS-DA-modelle) verdeel, wat van die maklikste klassifiseerbare klasse tot die moeilikste gevorder het. Die hiërargiese model was op die volledige NIR-spektrum (288 golfbande) gebaseer en is vir die hoofkategorieë met 'n algehele akkuraatheid van 93.3% uitgevoer. Die klassifikasie van gesonde witmielies (88.3%), pienk witmielies (83.3%) en geelmielies (75.0%) moet verkieslik verbeter word voordat die metode vir industrie-gradering geïmplementeer word. Pienk witmielies en geelmielies kan as gevolg van die teenwoordigheid van antosianien en beta-karoteen, respektiewelik, onderskei word wat beide 'n maksimum absorpsie in die sigbare gebied het en nie ‘n interaksie met NIR-bestraling het nie. Die gebruik van 'n spektrale beeldstelsel, insluitend die sigbare streek, sal na verwagting die klassifikasie van hierdie klasse verbeter. Na die bemoedigende sukses van mielie-gradering deur die volledige NIR-spektrum te gebruik, is golfbandvermindering en -optimalisering gedoen om 'n vereenvoudigde, maar noukeurige gradering van witmielies te bewerkstellig met behulp van 'n herberekende hiërargiese besluitneming. Drie golfbandseleksietegnieke is aangewend, naamlik golfbandvensters (48 golfbande), veranderlike belang in projeksie (VIP) (21 golfbande) en seleksie van kovariansie (CovSel) (13 golfbande). Daar was in al drie die golfbandmodelle 'n verlies aan prestasie. Die golfbandvensters (87.1% van die hoofkategorie-klassifikasie- akkuraatheid) en VIP-golfbandstelle (84,5% van die hoofkategorie-klassifikasie-akkuraatheid) is uitgevoer met soortgelyke klassifikasie-akkuraatheid in die verskillende kategorieë, maar die VIP-golfbandstel het minder as die helfte van die spektrale veranderlikes gebruik. Die CovSel-golfbandstel het die minste golfbande gebruik, maar het 'n onaanvaarbare verlies aan klassifikasie-akkuraatheid getoon (81.9% in die hoofkategorie- klassifikasie-akkuraatheid). In die algemeen is die VIP-golfbandstel (964, 1127, 1159, 1323, 1356, 1388, 1421, 1716, 1847, 1879, 1912, 1945, 2043, 2239, 2272, 2305, 2337, 2403, 2435, 2468 en 2501 nm), wat op slegs 7,3% van die 288 oorspronklike spektrale veranderlikes gebaseer was, is aanbeveel as die beste inruiling tussen instrumentprestasie en die verwagte koste van die stelsel. 'n Tweede kwessie in die Suid-Afrikaanse mieliebedryf is geadresseer, naamlik die opsporing van enkele asimptomaties Fusarium-besmette pitte. NIR hiperspektrale beelde van 224 sigbare (gesonde) pitte is verkry voor die pitte, in individuele steriele houers, ontkieming ondergaan het. Ontkieming het veroorsaak dat interne Fusarium-infeksies sigbaar identifiseerbaar word as eksterne swamgroei, wat later deur konvensionele mikrobiese toetsing bevestig is. Alhoewel slegs 3.3% van die pitte in die massamonsters sigbare verrottingsimptome vertoon het (wat tydens visuele inspeksie gemerk is), was 32.1% van die ontkiemde pitte asimptomaties besmet en is in staat om skadelike fumonisien-mikotoksiene te produseer. Sommige van hierdie grootmaatmonsters bevat fumonisienvlakke van 8 dpm (dubbel die wettige limiet), maar sou op grond van handmatige inspeksiemetodes veilig vir menslike verbruik verklaar word. Hierdie gebrek aan korrelasie tussen sigbare simptome en veiligheid beklemtoon die behoefte aan addisionele analitiese metodes om Fusarium- verwante risiko's te bepaal. Die voor-ontkiemende spektrale beelde van die onbesmette en asimptomaties besmette pitte is in twee klasse verdeel, en 'n PLS-DA-model het die mieliepitte met 'n klassifikasie- akkuraatheid van 67.0% geklassifiseer. Met inagneming van die hoë voedselveiligheidsrisiko verbonde aan fumonisiene, is NIR hiperspektrale beelding nie 'n uitvoerbare metode om asimptomatiese Fusarium-infeksies tydens die verwerking van Suid-Afrikaanse witmielies op te spoor nie. Die resultate van hierdie studie het getoon dat NIR hiper- en multispektrale beelding belowende analitiese tegnieke vir outomatiese mielie gradering is, maar nie vir die opsporing van asimptomatiese Fusarium-infeksie nie. Die resultate van die Fusarium-ontkiemingsstudie het egter insig gegee in die status van Fusarium-infeksie en fumonisiene in die Suid-Afrikaanse mieliebedryf, wat nog nie in die literatuur verskyn het nie, en het die behoefte aan industrie-vriendelike mikotoksien-toetsmetodes beklemtoon.

Please refer to this item in SUNScholar by using the following persistent URL: http://hdl.handle.net/10019.1/108366
This item appears in the following collections: