Plankton classification with high-throughput submersible holographic microscopy and transfer learning

Area of Science:

• Marine biology
• Microbial ecology
• Oceanography

Background:

• Plankton are crucial for marine food webs and ocean health indicators.
• Automating plankton classification from large image datasets presents ecological and technical challenges.
• Limited coupling of imaging instruments with classification algorithms hinders field data analysis.

Purpose of the Study:

• To develop and validate a high-throughput method for plankton classification using digital holographic microscopy and deep learning.
• To establish a baseline for classifying holographic plankton images, including rare and abundant species.

Main Methods:

• Utilized a portable digital in-line holographic microscope (The HoloSea) with 1.5 μm resolution.
• Employed intensity-based object detection within a volume for image processing.
• Applied four pre-trained convolutional neural networks for classifying over 3800 micro-mesoplankton images across 19 classes.

Main Results:

• Achieved high classifier performance (F1-scores >89%) quickly during training for each convolutional neural network.
• Demonstrated strong performance of off-the-shelf classifiers across all decision thresholds for most plankton classes.
• Successfully classified holographic images of both rare and plentiful plankton, including dinoflagellates and diatoms.

Conclusions:

• The study provides a compelling baseline for classifying holographic plankton images.
• Highlights the potential of deployable holographic microscopes for sampling diverse microbial eukaryotic communities.
• Supports the use of this technology for high-throughput plankton monitoring in marine environments.