Classification of walnut varieties with deep learning algorithms

Walnut is an important agricultural product with high economic value on a global scale. The accurate classification of walnut species is of great importance for research, conservation and quality control processes. However, traditional classification methods largely rely on manual processes, which are time-consuming and prone to human error. With technological advancements in the agricultural sector, the use of deep learning algorithms in the classification of products such as fruits, vegetables and grains is becoming increasingly widespread. In this thesis study, it is aimed to automatically classify walnut species by using machine learning algorithms and pre-trained deep learning models. Thus it is intended to reduce time loss, workload and misclassification rates. For this purpose, a special dataset consisting of a total of 2540 images belonging to the Chandler, Kaman1, Fernor, Yalova3 and Maras18 species was created. For the classification process, firstly, 4096 features were extracted from the dataset using the VGG16 deep learning model. These features were given as input to machine learning algorithms such as SVM, k-NN and Random Forest and the classification process was carried out. Afterwards, commonly used pre-trained deep learning architectures such as VGG16, VGG19, ResNet-50, DenseNet-121 and Xception were trained on the created original dataset using three different strategies: without data augmentation, with data augmentation and with a weighted loss function. According to the experimental results, classical machine learning algorithms presented lower accuracy values and among these algorithms, SVM showed the best performance with 90.10% accuracy. In contrast, the ResNet-50 model achieved the highest performance with an accuracy of 97.95% on the dataset with applied data augmentation. The Xception model obtained the best results with 98.54% accuracy when using a weighted loss function and 98.27% accuracy on the original dataset. The results obtained reveal the effectiveness of the ResNet-50 and Xception models in the automatic classification of walnut species and also show that deep learning technologies offer faster, more accurate and less labor-intensive alternatives compared to traditional methods. In general, this study contributes to the integration of deep learning technologies into agricultural production processes and proposes solutions to improve the efficiency of traditional classification methods.