Convolutional neural network application on a new middle Eocene radiolarian dataset

A new radiolarian image database was used to train a Convolutional Neural Network (CNN) for automatic image classification. The focus was on 39 commonly occurring nassellarian species, which are important for biostratigraphy.

The database consisted of tropical radiolarian assemblages from 129 middle Eocene samples retrieved from ODP Holes 1258A, 1259A, and 1260A (Demerara Rise). A total of 116 taxonomic classes were established, with 96 classes used for training a ResNet50 CNN. To represent the diverse radiolarian assemblage, some classes were formed by grouping forms based on external morphological criteria. This approach resulted in an 86.6% training accuracy.

A test set of 800 images from new samples obtained from Hole 1260A was used to validate the CNN, achieving a 75.69% accuracy. The focus then shifted to 39 well-known nassellarian species, using a total of 15,932 images from the new samples. The goal was to determine if the targeted species were correctly classified and explore potential real-world applications of the trained CNN.

Different prediction threshold values were experimented with. In most cases, a lower threshold value was preferred to ensure that all species were captured in the correct groups, even if it resulted in lower accuracies within the classes.