Aprendizado contrastivo para a classificação de espécimes de herbário da família Piperaceae

Abstract

Herbaria serve the function of documenting plant samples collected from nature, preserving them as exsiccatae: dried specimens mounted on cardstock, accompanied by detailed descriptions. These specimens are essential for taxonomic, biogeographical, and ecological studies, but they require proper identification. Typically, identification constitutes a bottleneck in the herbarium workflow, as it is an arduous task, prone to errors, and highly dependent on the capacity and availability of specialists. Recently, initiatives to mitigate this problem have leveraged the images available in virtual herbaria to create and release datasets. These datasets are intended for training predictive models through Machine Learning, aiming to support experts in the identification task. However, the performance of these models has been unsatisfactory for many botanical families due to factors such as high inter-species similarity, significant intra-species variability, and, primarily, the long-tailed distribution of the datasets, in which many species are represented by few samples. This highly challenging scenario is exemplified by the dataset of the Piperaceae Giseke family, where the performance achieved is still insufficient to effectively assist specialists. To address this issue, this research aims to investigate the use of Contrastive Learning to enhance the automated identification of species in herbaria, with a specific focus on the species of the Piperaceae family. In general, Contrastive Learning aims to structure a representation space by pulling similar samples closer while simultaneously pushing dissimilar ones apart. To evaluate this approach, a series of experiments is proposed using samples from the botanical family Piperaceae Giseke. The contrastive approach, specifically SimCLRv2 framework, demonstrated strong performance, achieving a peak F1-Score of 0.62. This result surpasses the traditional baseline by 26 percentage points. Moreover, the findings highlight the data efficiency of Contrastive Learning, it surpassed the fully-trained traditional model’s performance while using only 15% of the labeled data. These results validate Contrastive Learning as a robust and high-impact strategy for automated species classification.