Accurate classification of grapevine leaf plays a pivotal role in plant species identification, vineyards and agricultural automation. Manual classification can be time consuming and it may be liable to inconsistency. This paper is research on how well three deep learning techniques: Transfer Learning (TL), Deep Convolutional Neural Networks (DCNN) and Ensemble Learning can be used to automatically classify five types of leaves of grapevines: Ak, Ala_Idris, Buzgulu, Dimnit and Nazli. We apply three common architectures such as DenseNet201, InceptionV3 and VGG19 in the means and modes of TL and fine-tuned DCNN. Using the TL architecture, the DenseNet201, InceptionV3 and VGG19 attained an accuracy of 83%, 70% and 62% respectively. Trained DL models on a scale of Fine-tuned DCNN models have a viability that excelled considerably compared to TL models since DenseNet201 got 95%, InceptionV3 92%, and VGG19 84%. Also, a majority-vote ensemble that incorporated all the three architectures attained the best accuracy, which was 97.5%. The findings show that the ensemble approaches yield a better output when addressing this classification problem through the advantages associated with architectural diversity, and enhanced generalization. The paper makes a significant contribution by performing an in-depth assessment of model performance and offering effective guidance in the context of a production of automated and high-accuracy plant classification tools in a smart agricultural environment.