Phytoplankton detection and recognition in freshwater digital microscopy images using deep learning object detectors

Area of Science:

• Environmental Science
• Microbiology
• Computer Science

Background:

• Toxic phytoplankton pose risks to water quality, with toxins hard to remove by conventional methods.
• Manual analysis of phytoplankton is labor-intensive, subjective, and unreliable.
• Existing automated systems lack scalability and rely on classical image processing.

Purpose of the Study:

• To explore deep learning object detection for automated phytoplankton identification in microscopy images.
• To compare the performance of Faster R-CNN and RetinaNet for detecting toxic phytoplankton species.
• To develop a more reliable and scalable automated system for water quality monitoring.

Main Methods:

• Utilized a dataset of multi-specimen microscopy images captured with a systematic protocol.
• Implemented and evaluated two object detection models: Faster R-CNN (two-stage) and RetinaNet (one-stage).
• Focused on end-to-end trainable modules for integrated detection and recognition of phytoplankton.

Main Results:

• Faster R-CNN achieved high performance: 95.35% recall and 94.68% precision for W. naegeliana.
• Achieved 84.69% recall and 89.30% precision for A. spiroides, and 79.81% recall and 82.61% precision for D. sociale.
• The system demonstrated improved automation, abstraction, and workflow simplification compared to the state-of-the-art.

Conclusions:

• Deep learning, specifically Faster R-CNN, offers a robust solution for automated toxic phytoplankton detection.
• The developed system enhances reliability and scalability in water quality monitoring.
• Publicly releasing the dataset promotes reproducibility and further research in the field.