Multi-Label Benthic Foraminifera Identification With Convolutional Neural Networks

Fossil studies are crucial for understanding species evolution and uncovering the dynamic structure of the Earth. However, the examination and interpretation of fossil specimens is inherently complex and time-consuming, particularly when dealing with thin sections where microfossils and non-fossil structures often coexist. This study presents a comparative analysis of image classification models, specifically CNN, ResNet, VGG, DenseNet, and EfficientNet, aimed at enhancing the detection and classification of fossil specimens. We developed a custom Convolutional Neural Network (CNN) architecture tailored to the identification of benthic foraminifera using the Endless Forams dataset. The custom CNN achieved a training accuracy of 99% and a validation accuracy of 88%, indicating its robustness. ResNet-50 and VGG-16 models achieved average accuracy scores of 90% and 86%, respectively, demonstrating their comparative effectiveness. Furthermore, ResNet-50 and VGG-16 models were identified as particularly effective due to their advanced capabilities in handling high-dimensional data and capturing detailed image features. Our findings provide a comprehensive understanding of each model’s performance, supported by rigorous statistical evaluation, offering insights into their strengths and limitations within the domain of fossil image classification.