Identification of herbarium specimen sheet components from high-resolution images using deep learning.

Area of Science:

• Biodiversity informatics
• Computer vision
• Digital specimen imaging

Background:

• Herbarium specimens are crucial biodiversity data sources.
• Manual data extraction from digital images is time-consuming and error-prone.
• Advanced computer vision can automate and enhance data capture from herbarium images.

Purpose of the Study:

• To develop and evaluate an object detection model for identifying components on herbarium specimen sheets.
• To assess the model's generalisability across different herbaria.
• To provide a readily usable model for digitisation efforts.

Main Methods:

• Object detection model (YOLOv5) trained on 3371 annotated images from the University of Melbourne Herbarium (MELU).
• Model validation on 1000 annotated images, with performance metrics including precision, recall, and mean average precision (mAP).
• Generalisation testing on specimens from nine global herbaria and fine-tuning experiments.

Main Results:

• The MELU-trained 'sheet-component' model achieved high performance (mAP0.5-0.95 of 0.847), accurately identifying 11 component types.
• Specific labels like 'institutional' and 'annotation' showed strong prediction accuracy (mAP0.5-0.95 of 0.970 and 0.878).
• The model demonstrated generalisability across diverse herbaria (mAP0.5-0.95 between 0.68-0.89) and could be fine-tuned with minimal data.

Conclusions:

• Object detection models, like the YOLOv5-based 'sheet-component' model, significantly enhance biodiversity data extraction from herbarium images.
• The developed model is accurate, generalisable, and adaptable to new collections with limited retraining.
• Making the trained model weights available can support resource-constrained herbaria in their digitisation workflows.