Scaling down annotation needs: The capacity of self-supervised learning on diatom classification.

Area of Science:

• Life sciences
• Environmental monitoring
• Computational biology

Background:

• Diatoms are crucial biomarkers for environmental health assessment.
• Deep learning has advanced diatom classification, but supervised methods require extensive expert-annotated data.
• Scarce annotation data presents a significant challenge in diatom identification.

Purpose of the Study:

• To introduce self-supervised learning for diatom classification, addressing the challenge of limited annotated data.
• To evaluate the effectiveness of self-supervised pre-trained models in enhancing data utilization.
• To reduce the dependency on taxonomic experts in diatom identification.

Main Methods:

• Utilized self-supervised learning for pre-training diatom classification models.
• Fine-tuned pre-trained models with small labeled datasets.
• Investigated the impact of extended pre-training phases on annotation dependency.

Main Results:

• Self-supervised pre-trained models significantly improve the effectiveness of limited annotated data, especially for smaller datasets.
• Fine-tuning with only 50 samples per class achieved accuracy comparable to full supervised methods.
• Extended pre-training (1600 epochs) enabled comparable accuracy with just 30 samples per class, further reducing annotation needs.

Conclusions:

• Self-supervised learning offers a powerful solution for diatom classification with minimal annotated data.
• This methodology substantially decreases the reliance on highly skilled taxonomic experts.
• The findings pave the way for more accessible and efficient environmental biomonitoring using diatoms.