First automatic size measurements for the separation of dwarf birch and tree birch pollen in MIS 6 to MIS 1 records from Northern Germany

Martin Theuerkauf ¹, Elias Nehring ², Alexander Gillert ³

⁰Institute of Ecology Leuphana University Lüneburg Lüneburg Germany.

Ecology and Evolution +

|

June 17, 2024

View abstract on PubMed

Summary

We developed an automated method using machine learning to distinguish dwarf birch from tree birch pollen by measuring grain size. This technique rapidly and accurately quantifies past vegetation changes, improving our understanding of ancient environments.

Area Of Science

Paleobotany
Computational Biology
Image Analysis

Background

Distinguishing dwarf birch from tree birch pollen is crucial for reconstructing past vegetation in Northern Eurasia and North America.
Traditional pollen analysis faces challenges due to similar pollen morphology between birch species.
Pollen grain size statistics offer an objective method but are time-consuming.

Purpose Of The Study

To develop an automated approach for measuring birch pollen grain size using image recognition and machine learning.
To enable rapid and accurate differentiation of dwarf birch and tree birch pollen.
To enhance the quantification of past vegetation composition.

Main Methods

Application of the TOFSI algorithm for pollen detection and classification.
Utilizing a Resnet-18 neural network to select suitable birch pollen for measurement.
Employing semantic segmentation to determine pollen outline, area, and mean width.

Main Results

The developed method automates birch pollen size measurement (>1000 grains per hour) with high accuracy.
Successful application demonstrated on Lateglacial-Early Holocene and Mid to Late Pleistocene pollen records.
The tool facilitates rapid and regular size measurements of birch pollen.

Conclusions

This novel network-based tool significantly improves the efficiency and accuracy of birch pollen size analysis.
It enables more regular and extensive measurements, aiding in the understanding of birch species distribution and environmental responses.
Facilitates improved quantification of dwarf birch versus tree birch in paleoenvironmental reconstructions.

Keywords:

Holocene TOFSI automatic pollen recognition convolutional neural networks dwarf birch machine learning middle and upper Pleistocene tree birch