Humans in the loop: Community science and machine learning synergies for overcoming herbarium digitization...

Area of Science:

• Biodiversity Informatics
• Digital Humanities
• Computational Biology

Background:

• Digitizing natural history collections, particularly herbarium specimens, faces significant bottlenecks in converting imaged labels to digital text.
• Existing methods for transcribing specimen labels are often slow, labor-intensive, and prone to errors, hindering large-scale data accessibility.

Purpose of the Study:

• To develop and present a semi-automated solution for efficiently converting imaged herbarium labels into digital text.
• To leverage community science and machine learning to overcome the limitations of manual label transcription.

Main Methods:

• Development of a label finder and classifier using a humans-in-the-loop process with the Notes from Nature community science platform for training data.
• Implementation of an optimized optical character recognition (OCR) pipeline tailored for specimen labels, incorporating pre-processing, multiple OCR engines, and post-processing alignment techniques.

Main Results:

• Achieved over 93% success rate in finding and classifying main herbarium labels.
• The OCR pipeline demonstrated a greater than four-fold reduction in errors compared to standard open-source solutions.
• Integrated human validation through a custom Notes from Nature tool within the OCR workflow.

Conclusions:

• A functional set of tools, including a freely accessible web application, has been developed for herbarium digitization.
• The presented approach effectively addresses the efficiency bottleneck in converting imaged labels to digital text.
• Further integration of these services into existing toolkits can enhance broad community adoption and accelerate biodiversity informatics efforts.