Deep Learning-Based Classification Consisting of Pre-Trained Models and Proposed Model Using K-Fold Cross-Validation for Pistachio Species

Pistachio is a nut originating from the Middle East, and the main varieties grown and exported in Turkey are Kirmizi and Siirt pistachios. Due to their strategic importance in the agricultural economy, they need to be classified correctly. This study aims to classify Kirmizi and Siirt pistachios using various deep learning-based models and k-fold cross-validation. For this purpose, the seven convolutional neural network models trained by transfer learning and the proposed MSU-CNN model are used for classification with k-fold cross-validation. The dataset used in this study consists of 2148 images, 1232 of which belong to Kirmizi pistachio and 916 to Siirt pistachio. The K-fold cross-validation method is applied to enhance the generalization ability of the classification model, prevent overfitting, and improve performance reliability. The AlexNet, GoogLeNet, proposed MSU-CNN, VGG16, EfficientNet-b0, ResNet-18, Inception-v3, and ResNet-50 models achieved classification accuracies of 94.88%, 96.79%, 96.79%, 97.90%, 98.88%, 99.02%, 99.21%, and 99.63%, respectively, with average results based on 5-fold cross-validation and the highest accuracy attained by ResNet-50. The performance of models was evaluated using classification accuracy, sensitivity, specificity, precision, F1-score, and ROC-AUC values. According to the results, many of the proposed models proved to be effective in the identification of pistachio species.