The goal of this master thesis was to develop a mathematical model that based on hyperspectral imaging in the near infrared region could characterize wheat kernels from ten different sorts for three chosen quality parameters: relative hardness, protein content and vitreousness. The technology of hyperspectral imaging combines the best of imaging and spectroscopic technologies. It enables the prediction of quality parameters on individual objects internally and non-destructively. The vision for this technology is to use it as a new state-of-the-art method for in-line rapid separation and quality control of products from the agricultural industries. Therefore, it is of upmost importance to test the prediction ability of this technology.
Wheat was chosen as the object of interest because of the variation from grain to grain and an interest from the project partner CSORT to explore the possibilities of this technology on Russian wheat. The wheat was provided by CSORT together with additional parameters of them such as hardness class and Kjeldahl protein. The parameters were chosen based on available literature, their importance in grading and the possibility to measure all three parameters on the same kernels. Though preliminary experiments the reference methods for the parameter’s protein content, relative hardness and vitreousness were chosen as follows:
• Protein content: Flow injection analysis.
This reference method was chosen because of stability and timesaving compared to other available methods. The preliminary experiments showcased a systematic underprediction to Kjeldahl.
• Relative hardness: Rupture force measurement by simple compression.
This reference method was chosen because of being the only method to measure relative hardness on wheat kernels without being fully destructive. Preliminary experiments showcased that many external factors influenced the measurements such as weight and orientation.
• Vitreousness: Visual evaluation.
Approved as a standard method. Specialised equipment was not available to measure this parameter properly.
150 kernels were randomly and equally chosen from the ten sorts. 100 were to be used for model calibration and 50 as a test set. Under Image acquisition 142 wavelengths in the NIR of 938–1600 were available with spatial resolution of 2251x320. Four orientations were chosen.
The main experiments showcased that there was good prediction regarding protein content giving models with R2 for cross-validation as high as 0.775. The relative hardness prediction models were completely random and could not predict this parameter. It is suggested that the external factors critically influenced this parameter making the reference method unsuitable for relative hardness prediction. Regarding virtuousness, after relative hardness measurements many of the kernels were too damaged to visually estimate the vitreousness class. This gave okay to bad prediction models with a misclassification between non-vitreous and fully vitreous on 16.2 % for the test set. However, classification based on wheat hardness class gave 100% correct classification.
Literature has showcased that better predictions on hardness and vitreousness can be made with other reference methods. Therefore, it is suggested to predict two quality parameters instead with proper reference methods for better prediction.