This conference paper (purchase required) describes the development of a successful non-destructive test to discriminate 20 varieties of Indian wheat as either hard or soft wheat. The authors used near-infrared (NIR) hyperspectral imaging (spectral range 900–1700 nm) on a reference set of authentic samples. Data images were taken from both sides of the seed (ventral and dorsal side). The dataset included images of 20,160 seeds. The authors compared results from 5 different machine learning models. The models were trained using the mean spectral values extracted from the hyperspectral images. They also tried 5 different pre-processing techniques. They evaluated each model’s performance for both raw and preprocessed data. They found that the optimum model achieved a classification accuracy rate of 95.01% for amalgamated data (encompassing both ventral and dorsal side data), 95.05% for exclusively ventral side data, and 95.37% for exclusively dorsal side data.
Photo by Craig Manners on Unsplash
Comments