A predictive system for the classification of cereal species and varieties

• Developed a method using Convolutional Neural Networks to automate the identification of cereal grains. • Collected 13.854 high-resolution grain images across four cereal species and six commercial varieties for model training and evaluation. • Evaluated three CNN architectures (VGG16, ResNet50, and Xception) to benchmark classification performance. • ResNet50 and Xception models achieved 97% test accuracy on variety level. Cereals are fundamental to global food supply, and their seeds are the raw materials for many food industries. Grain industry testing and certification workflows include classification of cereal species. Traditional identification relies on expert’s knowledge and is conducted through manual visual assessment, a time-consuming process which can often be subjective. This study introduces a deep learning-based method utilizing Convolutional Neural Networks (CNNs), specifically VGG16, ResNet50, and Xception, to automate cereal classification. Using a digital camera, we captured cereal grain images from four different species (soft wheat, durum wheat, barley, oat) and six varieties (Soft Wheat: Sellas, Durum Wheat: Levante, Barley: Triptolemos and Dimitra, Oat: Flega and Pigasos) resulting in 13.854 image samples for model training and evaluation. The CNN models achieved test accuracies of 65%, 97% and 97%, for VGG16, ResNet50 and Xception respectively. Our findings demonstrate that CNN-based classification offers a reliable and efficient alternative to manual grain inspection, accurately distinguishing between different species and varieties. The outcomes of the proposed approach are both accurate and dependable, encouraging its practical implementation.