A comparative deep learning framework for grape leaf disease classification using EfficientNetB0, InceptionV3, and Xception

Detection of grapevine diseases should be done accurately and in a timely manner to maintain the health of the crop and the productivity of the vineyard. This paper focuses on the application of three well-known convolutional neural networks in the classification of grape leaf diseases and compares EfficientNetB0, InceptionV3, and Xception. These models were trained on a dataset of labeled images that included disease manifestations on the grapevine leaves, such as Black Rot, Isariopsis Leaf Spot, and Black Measles. The results of the experiments proved that out of all models compared, EfficientNetB0 is the best since the model’s validation accuracy was 99.5% and achieved macro AUC score of 1.00 on all classes, and on the test set it returned perfect scores which proved that the model was more robust and generalized better than others. The other two models that were compared to EfficientNet B0, Xception and Inception V3, performed slightly worse with a validation accuracy of 99% but remained strong and confident. Using confusion matrices and ROC curves for empirical evaluation also made it possible to see the difference of EfficientNet B0 to the other two models for classification accuracy and bias, where it outperformed the others in classifying diseases that have similar patterns. Most importantly, with the accuracy achieved, the lightweight and low computation requirement architecture of EfficientNet B0 is particularly useful for use in mobile real-time plant disease diagnostic systems. The results of this study serve as a valuable reference for selecting suitable models for automating grapevine disease detection, thereby further enhancing precision agriculture systems.