Identification of the cultivars of the wheat crop from their seed images using deep learning: convolutional neural networks

The characteristics and qualities of seeds (kernels) of wheat cultivars vary, in their size, shape and texture, genetic and biochemicals properties. Visual evaluations of color, size, shape and texture which are commonly used to identify cultivars of the crop seeds, are highly subjective and time-consuming. Although cultivar identification based on molecular biological properties of the seeds provide an edge over the traditional approaches, they are expensive and time-consuming. In recent years, computer vision technology is widely used for automation in agriculture. It makes a machine to see, hence instead of the human eye, this technology uses a camera and computer, to track, identify and measure the targets through image processing. The present study explores the possibilities of identifying cultivars of wheat seeds from their seed images using deep learning and computer vision techniques. Convolutional neural networks (CNN), were used in this study to categorize wheat seed images from ten different cultivars using deep learning technique. Five state of art CNN architectures, particularly VGG16, VGG19, ResNet50, ResNet101, and InceptionResNetV2, were compared based on nine thousand images captured from seeds of 10 wheat cultivars (HS 490, HD 2967, HI 1500, HB 208, C 306, CPAN 3004, DPW 621-50, HS 1097-17, HD 2864, AKAW 3722) in controlled illuminated chamber. Among the different CNN model tested in this study, InceptionResNetV2 was observed to be the best. Further, an augmented dataset created from the original seed image dataset were used to train and improve the model developed with the InceptionResNetV2. Further, the accuracy of the model to identify the wheat cultivars from their seed images was improved through hyperparameter optimization. The performance of the model in terms of its accuracy was enhanced from 68.3% with InceptionResNetV2 alone to 79.4% when combined with image augmentation and finally to 86.37% through hyper parameter optimisation with an average F1 score of 0.84. The results show that CNN based InceptionResNetV2 model could be advocated as a rapid, objective, non-destructive technique to identify the cultivars of wheat seeds from its images.