EfficientNetB0 Transfer Learning Improves Web-Based Grape Leaf Disease Classification

Abstract

Manual identification of grape leaf disease is often slow and inconsistent because several symptoms have similar color, lesion, and texture patterns. This study aimed to develop an accurate and efficient web-based classification system for grape leaf disease using transfer learning with the EfficientNetB0 architecture in a Convolutional Neural Network (CNN) framework. The research used a quantitative experimental method with 8,000 grape leaf images from four balanced classes: healthy, black rot, black measles, and leaf blight. Image data were first processed through resizing to 224 × 224 pixels, normalization, and augmentation through rotation, zooming, shifting, shearing, and horizontal flipping. The processed dataset was then divided into training, validation, and testing subsets using an 80:10:10 ratio. EfficientNetB0 was used as a pre-trained feature extractor, followed by additional classification layers and fine-tuning. Model performance was evaluated using accuracy, precision, recall, F1-score, and a confusion matrix, then deployed in a Flask-based web application. The results of this research are a training accuracy of 98.47%, a validation accuracy of 99.37%, and a validation loss of 0.0331. Based on the confusion-matrix counts, the class-wise evaluation produced macro precision, macro recall, and macro F1-score of approximately 99.37%, with most errors occurring between disease classes that have visually similar symptoms. These results indicate that EfficientNetB0 transfer learning can support accurate, efficient, and practical early detection of grape leaf diseases through a user-accessible web system.