Automatic identification of harmful algae based on multiple convolutional neural networks and transfer learning.

Area of Science:

• Marine Biology
• Computational Science
• Environmental Science

Background:

• Traditional phytoplankton monitoring relies on expert identification, which is labor-intensive, costly, and impractical for large-scale application.
• Harmful phytoplankton blooms can cause severe ecological damage, red tides, and contaminate drinking water with toxins.
• Accurate and efficient monitoring is crucial for maintaining aquatic ecosystem health and public safety.

Purpose of the Study:

• To develop an automated system for classifying algae cell images and identifying harmful phytoplankton.
• To improve the accuracy and efficiency of phytoplankton monitoring using deep learning and transfer learning techniques.
• To reduce the reliance on manual identification by experienced personnel.

Main Methods:

• Utilized transfer learning by fine-tuning five Convolutional Neural Networks (CNNs) models: AlexNet, VGG16, GoogLeNet, ResNet50, and MobileNetV2.
• Trained models on a dataset comprising 11 common harmful and 31 harmless phytoplankton genera.
• Developed a novel harmful phytoplankton identification method by combining the recognition outputs of the five fine-tuned CNN models.

Main Results:

• Fine-tuning the CNN models improved average classification accuracy by 11.9% compared to models without fine-tuning.
• The proposed identification method achieved a recall rate of 98.0% for harmful phytoplankton.
• Transfer learning significantly enhanced the recognition performance for harmful phytoplankton detection.

Conclusions:

• The developed deep learning approach provides an effective and accurate method for automated harmful phytoplankton identification.
• This automated system greatly reduces the workload of professional personnel in phytoplankton monitoring.
• The findings support the application of AI in safeguarding aquatic environments and public health from harmful algal events.