An algorithm competition for automatic species identification from herbarium specimens.

Area of Science:

• Botany
• Computer Science
• Machine Learning

Background:

• Plant biodiversity faces threats, with many species remaining undescribed.
• Over 50% of undescribed species are potentially housed in herbaria, awaiting identification.
• Automatic species identification algorithms can expedite the discovery of new plant species.

Purpose of the Study:

• To foster the development of automatic species identification algorithms.
• To leverage machine learning for accelerating species discovery from herbarium collections.
• To address the challenge of identifying flowering plant species within the Melastomataceae family.

Main Methods:

• The Herbarium 2019 dataset was submitted to the Fine-Grained Visual Categorization (FGVC6) competition on Kaggle.
• Focus was placed on the Melastomataceae family, utilizing 46,469 imaged herbarium specimens across 683 species.
• The dataset included varying numbers of specimens per species, reflecting typical herbarium collection characteristics.

Main Results:

• The FGVC6 Herbarium 2019 Challenge attracted 22 teams and 254 submitted models.
• The top-performing algorithms achieved over 88% accuracy in Melastomataceae species identification.
• The competition facilitated rapid advancements in automated plant identification.

Conclusions:

• Fine-Grained Visual Categorization competitions offer a valuable platform for computer vision and machine learning experts to tackle species recognition challenges.
• The Herbarium 2019 Challenge successfully combined herbarium resources, taxonomic knowledge, and interdisciplinary collaboration.
• This initiative highlights the potential of integrating computational approaches with traditional botanical expertise for biodiversity research.