Implementation Of Hybrid EfficientNet V2 And Vision Transformer for Apple Leaf Diseases Classification

Abstract

The apple farming industry faces challenges in managing apple leaf diseases. Current manual detection methods have limitations in expertise variability, time required, potential delays in identification leading to disease spread, and difficulty distinguishing diseases with similar visual symptoms. This research aims to develop an accurate, efficient, and automated apple leaf disease classification system using a hybrid approach that combines EfficientNet V2 architecture and Vision Transformer. The main objectives are to improve disease detection accuracy, reduce computational requirements, facilitate more effective plant management, and support modern agricultural practices in the apple industry. This research uses a hybrid deep learning model that integrates EfficientNet V2 and Vision Transformer components. Experiments were conducted on an apple leaf disease dataset to evaluate model performance. Results show the effectiveness of this method in classifying apple leaf diseases, achieving 98.56% accuracy and an F1 score of 0.9856 on test data. The proposed model has 15.6 million parameters, lighter than the original EfficientNetV2S model with 20 million parameters. Training time was reduced to 6 minutes 32 seconds compared to the original EfficientNetV2S model that required 8 minutes 41 seconds for 5 epochs on the same dataset.