Integrating Machine Learning with Flow-Imaging Microscopy for Automated Monitoring of Algal Blooms.

Area of Science:

• Environmental Science
• Aquatic Ecology
• Water Quality Monitoring

Background:

• Real-time monitoring of freshwater phytoplankton is crucial for early detection of harmful algal blooms (HABs).
• Automated systems are needed for efficient response by water management agencies.
• Existing methods face challenges with autonomous imaging artifacts and novel object identification.

Purpose of the Study:

• To adapt an automated flow-imaging device (ARTiMiS) for real-time algal monitoring in natural freshwater systems.
• To develop an image processing pipeline addressing artifacts and novel object detection.
• To enhance the scalability of automated detection in dynamic aquatic environments.

Main Methods:

• Developed an image processing pipeline for the ARTiMiS device.
• Utilized a random forest model to identify out-of-focus particles (89% accuracy).
• Implemented a custom algorithm for background particle detection (>97% accuracy).
• Employed a convolutional neural network (CNN) for taxonomic classification (95% accuracy in closed-set).
• Tested classification with rejection methods to handle novel particles.

Main Results:

• The pipeline effectively addresses flow-imaging artifacts like out-of-focus and background objects.
• Random forest achieved 89% accuracy for out-of-focus particles.
• Background particle detection exceeded 97% accuracy.
• CNN achieved 95% accuracy for closed-set classification.
• Classification with rejection improved precision by flagging unknown classes.

Conclusions:

• The integrated pipeline offers a robust solution for real-time harmful algal bloom monitoring in freshwater.
• Advances mitigate challenges with novel particles, reducing manual oversight.
• The system enhances the scalability of automated detection in dynamic aquatic environments.