Microscopic image recognition of diatoms based on deep learning.

Area of Science:

• Algology
• Ecology
• Bioinformatics

Background:

• Diatoms are vital indicators of aquatic ecosystem health and historical environmental conditions.
• Traditional diatom identification methods (morphological taxonomy, molecular detection) are resource-intensive and have limitations.
• Accurate diatom identification is essential for ecological monitoring and paleoclimate research.

Purpose of the Study:

• To develop an efficient and accurate automated method for diatom identification using deep learning.
• To create and utilize an extensive, augmented dataset of diatom images for model training.
• To evaluate and compare the performance of various deep learning algorithms for diatom classification.

Main Methods:

• Developed a large diatom image dataset (49,843 images) from 1042 species via augmentation.
• Trained and compared deep learning models, focusing on network architecture (ResNet152), batch size, and data augmentation techniques.
• Implemented a hybrid approach combining model prediction with cosine similarity to improve accuracy for challenging identifications.

Main Results:

• The ResNet152 network demonstrated superior performance, achieving top-1 accuracy of 85.97% and top-5 accuracy of 95.26% in identifying 1042 diatom species.
• The combined model prediction and cosine similarity method further enhanced accuracy to 86.07%.
• The study successfully validated the effectiveness of deep learning for large-scale diatom image recognition.

Conclusions:

• Deep learning, particularly the ResNet152 architecture, offers a powerful and accurate solution for diatom identification.
• The developed AI model significantly improves upon traditional methods in terms of speed and cost-effectiveness.
• This research has direct applications in water quality assessment, biodiversity monitoring, and understanding past environmental changes.