Foraminiferal image classification based on convolutional neural network considering data augmentation optimization

Abstract

Foraminifera are of utmost importance in paleoclimate and marine ecosystem research, with accurate classification being equally vital. Convolutional Neural Networks (CNNs) can realize the automatic classification of foraminiferal images, but they usually rely on data augmentation to address the issue of data scarcity. Despite the widespread use of data augmentation methods, the impacts of various augmentation methods on the classification of foraminifera remain unclear. In this study, we systematically evaluated the effects of different data augmentation methods on the classification performance of CNNs using three publicly available datasets. Experiments based on the ResNet-50 architecture showed that random rotation (RR), random flipping (RF), and random erasing (RE, ratio = 0.2) significantly improved the classification accuracy. The combined model of these three methods achieved accuracies of 89.4 %, 89.7 %, and 95.7 %, and F1 scores of 72.7 %, 72.8 %, and 84.3 % in the three tasks respectively. Compared with the basic model, the accuracy (A) increased by an average of 3.2 %, and the F1 score (F1) increased by an average of 7.1 %. This study confirms that selecting and combining appropriate data augmentation methods can effectively enhance the performance of foraminiferal image classification, with the combination of RR, RF, and RE being the most effective.