Foraminifera are single-celled marine organisms whose shells preserve valuable information about past ocean conditions, including temperature, salinity, and oxygen levels, but identifying them manually is slow and labor-intensive. The objective of this study was to evaluate whether a machine-learning model trained using Particle Trieur could automate the identification and classification of foraminifera species to improve efficiency and accuracy in research.

Foraminifera samples were collected and prepared, and 300 individual specimens were manually picked and identified. Each specimen was placed in a custom 3D-printed 50-slot microscope slide and imaged from multiple angles. Images were labeled using a standardized naming convention and uploaded to Particle Trieur, ensuring a minimum of 300 training images per species. The dataset was used to train a convolutional neural network based on the EfficientNetB4 architecture to assess model learning behavior and classification performance.

Results showed a downward trend in the loss function curve, indicating effective feature learning by the model. The confusion matrix produced an overall classification accuracy of 86.0%, and t-SNE visualization revealed distinct clustering patterns among species, suggesting that the model successfully identified morphological patterns within the dataset despite some overlap.

These findings demonstrate that Particle Trieur can accurately classify foraminifera species and has the potential to meet research-level reliability standards. Automating foraminifera identification could significantly improve the speed of environmental monitoring and paleoceanographic research, particularly in detecting hypoxia-related indicator genera such as Ammonia and Elphidium, enabling faster identification of emerging ocean dead zones.

 For more information: mekrischer@eckerd.edu