PAMUKKALE UNIVERSITY JOURNAL OF ENGINEERING SCIENCES-PAMUKKALE UNIVERSITESI MUHENDISLIK BILIMLERI DERGISI, vol.27, no.5, pp.618-626, 2021 (ESCI)
This paper presents a smart machine learning system for classification of hyperspectral wheat data based on deep learning methodology. For this purpose, the performances of AlexNet and VGG16 models were investigated for the classification of hyperspectral wheat samples. In this study, the Support Vector Machine (SVM) and Softmax classifiers were carried out to predict labels of wheat kernels. In order to evaluate the system performance, a new hyperspectral wheat test dataset was constructed using Visible-Near Infrared images (VNIR) including 50 wheat species with 220 images per specimen, as 11000 samples in total. With experiments applied on newly created test dataset, overall approximated accuracy rates of 96.00% and 99.00% determined by linear SVM classifier, in case of fully connected layer (FC6 and FC7) features for AlexNet and VGG16, respectively. From the Softmax predictions, the 92% and 70% of samples were correctly discriminated based on trained VGG16 and AlexNet models, respectively. The obtained superior results show that using a deep Convolutional Neural Networks (CNN) architecture is more efficient by the means of accurate discrimination of wheat species. The proposed deep learning based categorization system promises high accuracy results for the quality analysis, classification and disease detection in food.