Annotation-free learning of plankton for classification and anomaly detection.

Area of Science:

• Marine Biology
• Computational Biology
• Environmental Science

Background:

• Large plankton digital image datasets necessitate automated recognition and classification methods.
• Manual annotation and database representation hinder the use of machine learning for plankton taxonomy.
• Increasing data acquisition speed and size create bottlenecks in plankton image analysis.

Purpose of the Study:

• To develop novel, minimally supervised algorithms for accurate plankton species detection and classification.
• To address the limitations of manual data processing in large-scale plankton studies.
• To introduce an anomaly detection algorithm for unclassified plankton samples.

Main Methods:

• Development of a novel set of algorithms for plankton detection and classification with minimal supervision.
• Testing algorithms on a custom-built lensless digital device plankton dataset.
• Validation on a larger image dataset from the Woods Hole Oceanographic Institution.
• Introduction of a new algorithm for anomaly detection in unclassified samples.

Main Results:

• The proposed algorithms achieve accurate detection and classification of plankton species.
• Performance of the algorithms approaches that of existing supervised machine learning methods.
• Similar results were obtained on both custom and Woods Hole Oceanographic Institution datasets.
• The anomaly detection algorithm effectively identifies significant deviations from established classifications.

Conclusions:

• The developed algorithms offer an efficient and accurate solution for plankton image analysis.
• These methods reduce the reliance on manual annotation, overcoming data processing bottlenecks.
• The algorithms provide a new approach for rapid online environmental monitoring using intelligent detectors.