Using Hyperspectral imaging for characterisation of wheat kernels.
Translated title
Brug af hyperspektral afbilding for karakterisation af hvede kerner.
Author
Hansper, Frank Manfred
Term
4. term
Education
Publication year
2019
Submitted on
2019-06-07
Pages
127
Abstract
Formålet med denne kandidatopgave var at udvikle en matematisk model, der ved hjælp af hyperspektral billeddannelse i det nær-infrarøde (NIR) område kan karakterisere hvedekerner fra ti sorter efter tre kvalitetsparametre: relativ hårdhed (hvor hård kernen er ved tryk), proteindhold og vitreøsitet (hvor glasagtig kernen er indvendigt). Hyperspektral billeddannelse kombinerer kamera og spektrometer, så man kan måle mange bølgelængder på én gang og undersøge den enkelte kerne internt uden at ødelægge den. Visionen er at bruge teknologien til hurtig, in-line sortering og kvalitetskontrol i landbrugsindustrien, og derfor er det vigtigt at teste dens forudsigelsesevne. Vi valgte hvede, fordi der er stor variation fra kerne til kerne, og fordi projektpartneren CSORT ønskede at afprøve teknologien på russisk hvede. CSORT leverede kornene samt oplysninger som hårdhedsklasse og Kjeldahl-protein. Parametrene blev valgt ud fra litteraturen, deres betydning for klassificering og muligheden for at måle alle tre på de samme kerner. Reference- (sandheds-) målinger blev fastlagt i forforsøg: Proteindhold blev målt med flow-injektionsanalyse (stabil og tidsbesparende), som viste en systematisk underestimering i forhold til Kjeldahl. Relativ hårdhed blev målt som brudkraft ved simpel kompression, den eneste tilgang der kunne måle relativ hårdhed på hele kerner; resultaterne blev dog påvirket af ydre forhold som kernens vægt og orientering. Vitreøsitet blev vurderet visuelt (en standardmetode), men specialudstyr var ikke tilgængeligt. I alt blev 150 kerner udvalgt tilfældigt og ligeligt fra de ti sorter; 100 blev brugt til modelkalibrering og 50 som test. Hver kerne blev optaget ved 142 bølgelængder mellem 938 og 1600 nm og fra fire orienteringer. Resultaterne viste, at proteindhold kunne forudsiges med god nøjagtighed, med krydsvalideret R2 op til 0,775. Modellerne for relativ hårdhed var derimod tilfældige og ubrugelige, sandsynligvis fordi reference-metoden var for følsom over for ydre faktorer. For vitreøsitet betød kompressionsmålingerne, at mange kerner efterfølgende var for beskadigede til pålidelig visuel klassificering, og modellerne var fra middel til dårlige med 16,2 % fejlklassifikation mellem ikke-vitreøse og helt vitreøse kerner i testdatasættet. Til gengæld gav klassifikation baseret på hvedens hårdhedsklasse 100 % korrekt klassifikation. Litteraturen peger på, at andre reference-metoder kan give bedre forudsigelser for hårdhed og vitreøsitet. Vi foreslår derfor at fokusere på to kvalitetsparametre og anvende egnede reference-metoder for at forbedre præcisionen.
The aim of this master’s thesis was to build a mathematical model that uses near-infrared (NIR) hyperspectral imaging to characterize wheat kernels from ten types according to three quality traits: relative hardness (how much force is needed to compress a kernel), protein content, and vitreousness (how glassy the kernel appears inside). Hyperspectral imaging combines a camera and a spectrometer, capturing many wavelengths at once to probe the interior of individual kernels without destroying them. The long-term vision is rapid, in-line sorting and quality control in the agricultural industry, so testing the predictive ability of this technology is essential. Wheat was chosen because kernels vary widely and because the project partner CSORT was interested in applying the technology to Russian wheat. CSORT supplied the samples along with information such as hardness class and Kjeldahl protein. The three traits were selected based on the literature, their importance for grading, and the possibility of measuring all of them on the same kernels. Reference (ground-truth) methods were set in preliminary experiments: Protein content was measured by flow injection analysis (stable and time-saving), which showed a systematic underestimation relative to Kjeldahl. Relative hardness was measured as rupture force by simple compression, the only approach that could assess relative hardness on intact kernels; however, results were influenced by external factors such as kernel weight and orientation. Vitreousness was assessed by visual inspection (a standard method), but specialized equipment was not available. In total, 150 kernels were randomly and evenly selected from the ten types; 100 were used for model calibration and 50 for testing. Each kernel was imaged at 142 wavelengths between 938 and 1600 nm and from four orientations. Results showed good predictions for protein, with cross-validated R2 values up to 0.775. In contrast, models for relative hardness were essentially random and could not predict this trait, likely because the reference method was too sensitive to external factors. For vitreousness, compression tests left many kernels too damaged for reliable visual grading, and models ranged from fair to poor, with 16.2% misclassification between non-vitreous and fully vitreous kernels in the test set. However, classification based on wheat hardness class achieved 100% correct classification. Prior studies indicate that other reference methods can yield better predictions for hardness and vitreousness. We therefore suggest focusing on two quality parameters and using appropriate reference methods to improve prediction performance.
[This abstract was generated with the help of AI]
Keywords
Documents