Classification of wheat varieties with image-based deep learning

Wheat is an important grain in the food chain. It is important in terms of efficiency and economy to use wheat in the appropriate area according to its varieties. Breeding studies make varieties of wheat physically similar to each other and make it difficult to classify according to variety. An image-based deep learning approach is proposed to classify wheat accurately and reduce classification difficulties. Twenty-four varieties of wheat were used in the study and these varieties were harvested in five provinces of Turkey. The reflectance values of the wheat varieties were measured with a near-infrared spectrometer device and the measured reflectance values were used to create wheat images with a suggested method. With this method, low-dimensional images were created with reflection data that take up less space instead of a high-resolution image and a high-storage space requirement. With the classification processes, a 96.55% accuracy was obtained for the hard-white wheat class, 98.70% for the hard-red wheat class and 99.52% for all wheat varieties. The results show that the proposed image generation method with reflection data and the deep learning model is sufficient in classification. This method offers a new approach to the classification of wheat-like cereals. The proposed method can be considered an alternative classification method in the wheat production and trading sectors. It can also be used in the industry by integrating it into hardware with low memory/low processing power. This scenario can also be considered as a method for classifying grain groups other than wheat.