Pollen identification through convolutional neural networks: First application on a full fossil pollen sequence.

Area of Science:

• Paleoecology
• Palynology
• Machine Learning

Background:

• Fossil pollen analysis aids in reconstructing past environments and vegetation.
• Automating fossil pollen identification can save time, reduce costs, and minimize bias.
• Current Convolutional Neural Network (CNN) models face challenges due to limited fossilized pollen data and a gap between fresh and fossil pollen training data.

Purpose of the Study:

• To develop and test a large-scale automated fossil pollen identification workflow.
• To bridge the gap between fresh and fossil pollen data for machine learning models.
• To evaluate the performance of CNNs trained on fresh pollen for fossilized pollen identification.

Main Methods:

• An accelerated fossil pollen extraction protocol was employed.
• Convolutional Neural Networks were trained on labeled fresh pollen from common Northeastern American species.
• The model was tested on both fresh pollen and a large dataset of 196,526 fossil pollen images.

Main Results:

• The model achieved 91.2% average per-class accuracy on fresh pollen.
• Performance decreased significantly on fossilized pollen, with overconfident predictions.
• Despite performance limitations, general abundance patterns in fossil data aligned with traditional palynologist identifications.

Conclusions:

• The proposed method serves as a proof of concept for automated fossil pollen identification.
• Current models require further refinement to accurately classify entire fossil pollen sequences.
• Bridging the data gap between fresh and fossil pollen is crucial for advancing automated palynology.