Early detection of walnut cultivars by deep learning method using shoot and leaf images

Our country is among the important countries of the world in walnut cultivation. The amount of walnut grown in our country is not at a level to meet the country's needs. The main reason why it is not sufficient is that a conscious walnut cultivation cannot be carried out in our country. In walnut cultivation with high nutritional value, it is important to choose the walnut species suitable for climate and soil conditions. In walnut cultivation, choosing the right walnut species should be the first thing to do, since the garden establishment is costly and the walnut fruiting period takes years. It is quite difficult to distinguish the species of walnut trees. Walnut species are very similar to each other in terms of colour, texture and leaf structure. In general, to determine the type of walnut, the tree is expected to bear fruit. In the literature, most of the studies carried out to identify walnut species are classification studies on walnut fruit. In this study, not only walnut leaf images were used but also walnut shoot images were used for the first time. As a contribution to the literature, data sets containing 3 552 shoot and 10 000 leaf images of 10 different walnut species were created. The classification of the images in the shoot and leaf datasets were performed separately with deep learning models and compared. At the same time, in order to evaluate the success rate, hybrid models were developed and classification studies were also carried out. The most successful deep learning model proposed for the classification of the shoot data set was the DenseNet201 model with a success rate of 98.52%. Again, 98.30% success rate was achieved with the Dense-Walnut model developed in the training of the shoot dataset with hybrid models. For the classification of the leaf dataset, the DenseNet201 model, which is the most successful deep learning model, achieved a success rate of 95.52%. In the classification study of hybrid models, DenseNet201+EfficientNetB3 model achieved 97.84% success rate on the leaf data set.