This work proposes an efficient approach for image classification in fine-granularity scenarios, using self-supervised representations generated by the DINOv2 model. Unlike traditional methods based on deep convolutional networks, our solution employs DINOv2 as a visual feature extractor, followed by classical machine learning classifiers. This approach significantly reduces computational cost and training time, without compromising performance. To validate our proposal, we use the ‘102 Category Flower Dataset’, characterized by a high diversity of classes with visual similarities between species. The results show that the Logistic Regression-based model achieves an accuracy of 99.71%, outperforming or equalling state-of-the-art methods. Furthermore, the advantages of using self-supervised representations in complex contexts are discussed and future lines of work are proposed to improve the specialization of the model in the plant domain.