Test metrics.

In the realm of plant classification, the classification of grape leaf varieties has long presented a complex challenge. Aiming to enhance the accuracy and generalization ability of grape leaf variety classification, this study proposes a novel approach that employs an optimized Densenet201 model for grape leaf classification. Initially, grape leaf images from five distinct varieties were meticulously collected to construct a comprehensive grape leaf dataset. To augment the diversity of the dataset, the parameters of data augmentation were delicately adjusted, with an increase in the rotation range, translation range, and so on. Subsequently, BatchNormalization and GlobalAveragePooling2D layers were incorporated to achieve feature normalization and pooling. Simultaneously, the parameters of the Dropout layer were optimized to effectively mitigate the issue of overfitting. Additionally, the number of neurons and layers in the Dense layer were varied to explore diverse network structures and pursue superior performance. Moreover, the parameters of the Adam optimizer were meticulously tuned to attain the optimal performance, and the model’s performance was further enhanced by extracting image features. The experimental results demonstrate that, in comparison with the densenet121, densenet169, resnet50, and densenet201 models, the optimized Densenet201 model showcases outstanding performance in grape leaf variety classification, remarkably improving the classification accuracy and generalization ability. This research provides a more efficient method for grape leaf variety classification.