Can Synthetic Plant Images From Generative Models Facilitate Rare Species Identification and Classification?

In the quest to bridge the gap between the burgeoning capabilities of text-to-image generative models and the pragmatic demands of botanical classification, our study delves into the untapped potential of synthetic images for identifying and differentiating rare plant species. By rigorously evaluating the efficacy of cutting-edge generative models, including open-sourced and proprietary frameworks, we illuminate the advantages and inherent challenges of employing synthetic data in zero-shot and few-shot learning scenarios. Our research demonstrates that the zero-shot method sees a marked improvement of 29% over the pre-trained weights, with an average increment of 12%. Furthermore, the few-shot method improves the performance by an additional 31%, with an average increment of 19%, achieving new state-of-the-art classification results on rare flora. Through a comprehensive analysis that spans diverse species and models, we unravel the complexities of synthetic data integration, proposing innovative strategies to harness its full potential for the conservation and study of botanical diversity. This investigation stands at the forefront of combining advanced machine learning techniques with environmental science, paving the way for new advancements in accurately identifying and preserving rare plant species.