Deep Learning Classification of Lake Zooplankton

Area of Science:

• Environmental science
• Aquatic ecology
• Computational biology

Background:

• Plankton are vital indicators of freshwater ecosystem health.
• Manual plankton identification is time-consuming and costly.
• Automated plankton imaging offers real-time monitoring but faces annotation challenges.

Purpose of the Study:

• Develop and optimize deep learning models for lake plankton identification.
• Address the challenge of manual annotation in large plankton image datasets.
• Provide operational strategies for accurate plankton classification.

Main Methods:

• Annotated over 17,900 images of zooplankton and phytoplankton from Lake Greifensee.
• Developed deep learning classifiers, focusing on transfer learning and ensembling.
• Utilized the Dual Scripps Plankton Camera for image acquisition.

Main Results:

• Achieved 98% accuracy and 93% F1 score in plankton image classification.
• Models outperformed existing methods on diverse automated plankton datasets.
• Developed models demonstrated high performance in identifying 35 plankton classes.

Conclusions:

• Deep learning models offer an efficient and accurate solution for lake plankton identification.
• The developed models and annotated data are valuable resources for researchers.
• This approach enhances the feasibility of large-scale, real-time plankton monitoring.