Intelligent classification of urban grassland herbs based on ResViT modeling

In the field of grassland monitoring in arid and semi-arid areas, the utility and contribution of the classification and recognition of herbaceous plants cannot be underestimated. However, current deep learning models continue to have shortcomings with respect to tasks involving substantial sample data and small scale. Urban grassland monitoring can effectively enable assessments of the growth status of grasslands and provide information for evaluating the potential harm to local ecosystems based on the classification of herbs. On the basis the ViT (Vision Transformer) and ResNet50 (Residual Network 50 layers) models, in this study, we constructed a hybrid neural network model referred to as ResViT, which is superior to the AlexNet, ResNet50, and VGG19 models in terms of test set accuracy, average recall rates, and F1 scores. ResViT can be trained within half the time needed for VGG19, and achieved an accuracy of 95.45% and an F1 score of 0.95 when used to perform a test set of 16 classification tasks. To summarize, the ResViT model can accurately and efficiently accomplish image recognition tasks for the classification of herbs and has distinct advantages compared with the AlexNet, ResNet50, and VGG19 models. It has shown excellent performance when used to assess heavily small-scale datasets, significantly reducing the cost of preliminary data preparation, whilst also improving training efficiency and reducing training time. Consequently, the establishment of ResViT offers a novel perspective for research in the field of herb classification, and it is anticipated that this model will play key roles in extensive applications for grassland monitoring in arid and semi-arid areas.