Deductive automated pollen classification in environmental samples via exploratory deep learning and imaging flow...

Area of Science:

• Environmental Science
• Computational Biology
• Paleontology

Background:

• Palynology, the study of pollen and spores, traditionally relies on manual microscopic analysis.
• Manual methods are time-consuming, limiting taxonomic precision and sampling frequency.
• This restricts the quality of data used in climate change and pollen forecasting models.

Purpose of the Study:

• To develop a flexible artificial intelligence (AI) network for automated pollen classification in environmental samples.
• To improve the accuracy and efficiency of pollen identification beyond traditional methods.
• To create a transferable tool for analyzing 'real-world' environmental samples.

Main Methods:

• Combined imaging flow cytometry with Guided Deep Learning (GDL).
• Applied the network to identify and categorize pollen grains from ~5500 Cal yr BP old lake sediments.
• Trained the network to discriminate pollen to the species level and classify unseen pollen to higher phylogenetic ranks.

Main Results:

• The AI network accurately classified known pollen to the species level.
• Unseen pollen samples were classified to the likely phylogenetic order, family, and genus.
• Achieved improved accuracy compared to pure deep learning techniques.

Conclusions:

• The developed approach offers a rapid, accurate, and transferable tool for exploratory pollen classification.
• This method enhances taxonomic resolution and provides a more detailed spatial and temporal understanding of environmental pollen.
• The AI-driven tool has the potential to significantly advance various aspects of palynological research and applications.