The Herbarium 2021 Half-Earth Challenge Dataset and Machine Learning Competition.

Area of Science:

• Botany
• Biodiversity Informatics
• Computer Vision

Background:

• Herbarium sheets are crucial for understanding botanical history, evolution, and biodiversity.
• Digitization of herbaria and advances in fine-grained visual classification offer opportunities for botanical research.
• Existing datasets for automatic taxon recognition are limited in size and diversity.

Purpose of the Study:

• To introduce a large-scale, diverse dataset for automatic taxon recognition from herbarium specimens.
• To facilitate research in botanical identification and biodiversity informatics.
• To encourage the development of advanced machine learning models for analyzing herbarium data.

Main Methods:

• Development of the Herbarium 2021 Half-Earth dataset, comprising diverse herbarium specimen images.
• Aggregation and alignment of taxa names to a common reference.
• Hosting the Herbarium 2021 Half-Earth challenge to spur model development.

Main Results:

• The Herbarium 2021 Half-Earth dataset is the largest and most diverse to date for automatic taxon recognition.
• The challenge successfully encouraged the development of models for identifying taxa from herbarium images.
• The dataset and challenge address limitations of previous resources in size and diversity.

Conclusions:

• The Herbarium 2021 Half-Earth dataset is a significant resource for advancing automated plant identification.
• This initiative supports broader research in botany, evolution, and biodiversity.
• Further development in fine-grained visual classification is crucial for herbaria digitization efforts.