Comparative analysis of CNN-based pre-trained architectures for grape leaf disease detection

Abstract-Addressing plant diseases is a critical concern in agriculture, posing a significant threat to food quality. Currently, the identification of crop diseases is done by the naked eye which is time-consuming and also error-prone.[1] To take timely action against plant diseases, it is essential to monitor and detect diseases as early as possible. To reduce identification time and human error, Machine learning models, specifically deep learning models are very beneficial.[1] The models used for predictions are based on Convolutional Neural Networks (CNN) which include ResNet50, VGG16, and EfficientNetV2B3. The study utilizes an image dataset featuring grape crops sourced from Grapevine Disease Images. The three most common grape leaf diseases—black rot, ESCA, and leaf blight—have been targeted due to their adverse impact on crop yield. The findings aim to determine the most effective model for accurate and early detection of grape leaf diseases, contributing to precision grape farming. Among the three models assessed, ResNet50 is the most effective, achieving the highest accuracy of 99.11% and equal values for precision, F1 score, and recall of 0.99. These results determine that ResNet50 is significantly reducing the time and error of identification, offering the advancement of precision agriculture in the grape farming sector. Keywords: Grape leaf disease detection, image processing, Convolutional Neural Network, Data augmentation, Black rot, ESCA, leaf blight, ResNet50, VGG16, EfficientNetV2B3.