Wheat and triticale whole grain near infrared hyperspectral imaging for protein, moisture and kernel hardness quantification

Orth, Sebastian Helmut (2021-03)

Thesis (MScFoodSc)--Stellenbosch University, 2021.

Thesis

ENGLISH ABSTRACT: Wheat (Triticum aestivum) is one of the most important cereal crops grown globally. Triticale (× Triticosecale sp. Wittmack ex A. Camus 1927) is an important cereal crop for feed and fodder production and is also emerging as an alternative cereal for human consumption. Both these cereals are grown and produced in a diverse climatic environment and they vary with regards to their physicochemical properties. Quantitative techniques for determining protein and moisture content and kernel hardness is of importance for grading of the grains. The use of non-invasive and rapid techniques such as near-infrared hyperspectral imaging (NIR-HSI) show potential for quantification of these quality parameters. This study aimed to investigate the use of NIR-HSI (HySpex SWIR 384) with partial least squares regression (PLS-R) analysis for wheat and triticale bulk sample and single kernel image approaches. The study considered South African wheat and triticale samples produced in three Western Cape localities, i.e. Napier, Tygerhoek and Vredenburg, comprising 180 wheat and 177 triticale samples. Of these, 39 kernels per sample were used for single kernel protein and moisture content and kernel hardness prediction, resulting in data sets with a total of 7020 wheat, 6903 triticale and 13923 combined single kernel images. This was further split into training (70%) and validation (30%) sets using the Duplex algorithm. NIR (1100-2100 nm) hyperspectral images were acquired and the spectral data obtained for each pixel were averaged for each kernel. PLS-R was used to develop quantitative prediction models. Principal component analysis (PCA) was performed on the average spectral data and the PCA plot (PC1 vs. PC2) indicated separation between locality, with both wheat and triticale separating in the direction of PC1 from left to right. A PCA (PC1 vs. PC2) was performed for the wheat and triticale combined data set – no separation was noted. Bulk sample protein, moisture content and kernel hardness models were first evaluated which showed favourable prediction accuracy, comparable to conventional NIR spectroscopy studies performed on wheat and triticale. The combined wheat and triticale data sets for protein and moisture content and kernel hardness prediction had RMSEP-values of 0.41%, 0.49% and 8.66, respectively. Single kernel analysis involved two main quantitative data analysis methods (PLS-R and Robust-PLS) which were tested with an independent test set. The results being favourable for the conventional PLS-R method when only the validation set RMSEP (protein content: 0.37-0.84%, moisture content: 0.23-0.57% and kernel hardness: 1.74-3.64) was considered. The independent test set for protein content prediction achieved better results with the Robust-PLS (RMSEP protein content: 1.95-2.37%) method, proving that the method did indeed have an effect on making the calibration data sets more robust. Spectral imaging showed that it is capable to accurately quantifying protein and moisture content and kernel hardness of bulk and single kernel samples – good robust models proved to optimally quantify these parameters. The technique shows good potential for further study and to build onto the current data sets in order to increase variance across seasons. Further the technique showcases the functionality of SK NIR-HSI analysis and can be used both as a quality control measure and as an early generation selection method by the grain breeding sector.

AFRIKAANSE OPSOMMING: Koring (Triticum aestivum) is een van die wêreld se belangrikste graan gewasse. Korog (× Triticosecale sp. Wittmack ex A. Camus 1927) is ʼn belangrike graan gewas vir aangeplante weiding en kuilvoer produksie en is ook ʼn opkomende alternatiewe graan vir menslike gebruik. Albei hierdie graan soorte word in ʼn diverse klimatologiese omgewing verbou en daar is ʼn groot variasie tussen grootmaat monsters en tussen enkel sade vanuit ʼn monster. Kwantitatiewe tegnieke om graan proteïen- en voginhoud en hardheid te bepaal is van belang vir die gradering daarvan. Die gebruik van nie-indringende en vinnige tegnieke soos naby infrarooi (NIR) hiperspektrale beelding wys potensiaal vir kwantifisering van kwaliteiteienskappe. Hierdie studie was daarop gemik om ondersoek in te stel tot die gebruik van NIR hiperspektrale (HySpex SWIR 384) beelding met parsiële kleinste kwadrate regressie as die data analise metode, vir koring en ook korog monsters op ʼn grootmaat monster asook ʼn enkel saad beelding benadering. Die studie het Suid-Afrikaanse koring en korog monsters oorweeg wat verbou is in drie distrikte in die Wes-Kaap provinsie naamlik Napier, Tygerhoek en Vredenburg, wat verder afgebreek is na 180 koring en 177 korog monsters. Vanaf die grootmaat monsters is 39 sade per monster gebruik vir enkel saad analise vir proteïen, vog en hardheid inhoud bepalings, wat ʼn totaal van 7020 koring, 6903 korog en ʼn gekombineerde 13923 sade opmaak vir elke data stel. NIR hiperspektrale beelding (1100-2100nm) is gebruik om pixel en daaropvolgende spektrale data te verkry vanaf die sade en parsiële kleinste kwadrate regressie is gebruik as die kwantitatiewe data analise metode. Hoofkomponent analise (HKA) vir HK1 teen HK2 is uitgeoefen vir die bepaling van skeiding tussen monsters gebaseer op verbouings lokaliteit. Beide koring en korog datastelle wys daarop dat daar skeiding oor HK1 is van links na regs. ʼn HKA (HK1 teen HK2) is ook toegepas op die kombinasie datastel vir koring en korog, dit het geen skeiding tussen die twee graan soorte getoon nie. Grootmaat proteïen, vog en korrel hardheid modelle is toegepas op koring en korog en het gewys op gunstige voorspellings akkuraatheid wat vergelykbaar is met studies wat gefokus het op die gebruik van konvensionele NIR spektroskopie op koring en korog. Die gekombineerde data stelle vir proteïen- en voginhoud en hardheid bepaling het ʼn gemiddelde vierkantswortel fout van voorspelling (GVFV) waardes van 0.41%, 0.49% en 8.66, onderskeidelik gehad. Vir enkel saad analise, is twee kwantitatiewe data analise metodes gebruik (parsiële kleinste kwadraat regressie en robuust parsiële kleinste kwadraat regressie) wat getoets is teenoor ʼn onafhanklike toets stel. Die resultate was gunstig vir die konvensionele parsiële kleinste kwadraat regressie metode wanneer slegs gekyk is na die GVFV van die validasie stel. Die onafhanklike toets stel vir proteïeninhoud bepaling het ʼn beter GVFV gehad vir die robuust parsiële kleinste kwadrate regressie en wys daarop dat die kalibrasie van die modelle meer robuuste voorspellings maak. Spektrale beelding het gewys dat dit ʼn akkurate metode is om proteïen- en voginhoud en hardheid van grootmaat sowel as enkel sade te bepaal. Met optimale resultate geskik vir meer robuuste modelle vir verdere kwantifisering van kalibrasie parameters. Die tegnieke wys potensiaal vir verdere studie en om verder te bou op die huidige data stelle vir meer variasie oor seisoene. Verder word die funksionaliteit van NIR hiperspektrale beelding uit gewys en die metode kan sy plek vind in kwaliteit beheer sowel as graan seleksie in die graan teler sektor.

Please refer to this item in SUNScholar by using the following persistent URL: http://hdl.handle.net/10019.1/110017
This item appears in the following collections: