Classification of wheat kernels with near-infrared hyperspectral imaging

Naude, Alexia Irene (2019-12)

Thesis (MScFoodSc)--Stellenbosch University, 2019.

Thesis

ENGLISH ABSTRACT: Wheat (Triticum spp.) is a widely grown cereal crop and is one of the most important staple foods internationally. In South Africa, it is the second most produced cereal where the majority is used for the production of bread. Grading is an important step in wheat production to ensure that the grains are of an acceptable quality for food processing and consumption. The kernels are broadly assessed in terms of cleanliness, ripeness, damage and foreign materials. Current wheat grading practices are manual, tedious, time-consuming, and subjective. More modern analytical methods that can rapidly and accurately assess wheat quality are continually addressed. In this study, four common visual defects namely heat-damaged, Fusarium-damaged, sprout-damaged, and immature wheat kernels were considered. The study aimed to investigate the feasibility of hyperspectral imaging and chemometrics to discriminate sound wheat from the four defective categories. Dichotomous and multiclass classifications were performed, where the objectives were to identify the optimal pre-processing technique and classification algorithm. For the classification between sound and heat-damaged wheat, the logistic regression classifier, applied to data pre-processed with a combination of 2nd derivatives and Standard Normal Variate corrected spectra, provided the highest classification accuracy (99.2%). For sprouting analysis, Support Vector Machines applied to 2nd derivative and Standard Normal Variate corrected spectra achieved the highest classification accuracy (98.6%). Fusariumdamaged wheat scored 100% classification accuracy with the Random Forests, decision trees or k-Nearest Neighbours classifier applied to the spectra pre-processed with 2nd derivatives. For immature wheat, the best classification was accomplished with 2nd derivatives and Partial Least Squares Discriminant Analysis, where 99.3% accuracy was achieved. The multiclass analyses had a decreased performance compared to the dichotomous analyses. The highest classification accuracy was achieved by pre-processing with Standard Normal Variate, de-trending and 2nd derivatives, and classifying with Support Vector Machines (C: 1000; γ: 1). This model attained an accuracy of 84.6%. Given that wheat is an agricultural product that varies significantly in inherent characteristics due to genetics, cultivation practices, handling and storage, the overall results achieved are highly successful. Hyperspectral imaging proved to be capable of effectively discriminating sound wheat from common occurring defects. This technology, therefore, has the potential to asses wheat quality, offering a rapid, accurate and objective alternative grading method to the cereal industry.

AFRIKAANSE OPSOMMING: Koring (Triticum spp.) is 'n wydverboude graangewas en is internasionaal een van die belangrikste stapelvoedsels. Dit is in Suid-Afrika die tweede meeste geproduseerde graan, en die meerderheid word vir die produksie van brood gebruik. Gradering is 'n belangrike stap in koringproduksie om te verseker dat die monster 'n aanvaarbare gehalte vir voedselverwerking en -verbruik is. Die monster word breedweg beoordeel in terme van netheid (skoon), rypheid (onryp), skade en vreemde materiale. Huidige koring gradeer praktyke is afhanklik van menslike oordeel en word per hand en oog gradeer wat tydrowend is en wat ook tot gebrek aan konsentrasie lei. Moderne analitiese metodes wat koring kwaliteit vinnig en akkuraat kan bepaal, word deurlopend ontwikkel. In hierdie studie is vier algemene visuele defekte, naamlik hittebeskadigde, Fusarium-beskadigde, uitgeloopte en onvolwasse koringpitte, oorweeg. Die studie het ten doel gehad om die uitvoerbaarheid van hiperspektrale beelding en chemometrie te ondersoek om gesonde koring uit die vier gebrekkige kategorieë te onderskei. Dichotome en multiklas klassifikasies is uitgevoer, met die doel om die optimale voorverwerkingstegniek en klassifikasie-algoritme te identifiseer. Vir die klassifikasie tussen gesonde en hittebeskadigde koring, het die logistieke regressie-klassifiseerder, toegepas op data wat vooraf verwerk was met 'n kombinasie van tweede orde afgeleides en “Standard Normal Variate” gekorrigeerde spektra, die hoogste klassifikasie akkuraatheid (99.2%) behaal. Vir ontkiemingsanalise, het “Support Vector Machines” toegepas op die tweede orde afgeleide en “Standard Normal Variate” gekorrigeerde spektra, die hoogste klassifikasie akkuraatheid (98.6%) behaal. Koring met Fusariumbeskadiging is 100% ge-identifiseer met “Random Forests”, “Decision Trees” of “k-nearest neighbours” wat toegepas was op die spektra wat vooraf met die tweede orde afgeleides verwerk is. Onvolwasse koring was die beste ge-identifiseer met die tweede orde afgeleide metode en “Partial Least Squares Discriminant Analysis”, waar 99.3% akkuraatheid behaal was. Die multiklas ontledingsmetode het swakker gevaar in vergelyking met die digotome ontledingsmetode. Die hoogste klassifikasie-akkuraatheid was behaal deur voorafverwerking met “Standard Normal Variate”, “de-trending” en tweede orde afgeleides, en te klassifiseer met “Support Vector Machines” (C: 1000; γ: 1). Hierdie model het 'n akkuraatheid van 84.6% behaal. Aangesien koring 'n landbouproduk is wat baie verskil in inherente eienskappe as gevolg van genetika, verbouingspraktyke, hantering en opberging, is die algehele resultate wat behaal is baie suksesvol. Hiperspektrale beelding het die vermoë om goeie gehalte koring effektief te onderskei van algemene koring defekte. Hierdie tegnologie het dus die potensiaal om koringkwaliteit te beoordeel en bied 'n vinnige, akkurate en objektiewe alternatiewe graderingsmetode vir die graanbedryf.

Please refer to this item in SUNScholar by using the following persistent URL: http://hdl.handle.net/10019.1/106961
This item appears in the following collections: