Advancing the automated foraminifera fossil identification through scanning electron microscopy image classification: A convolutional neural network approach

Abstract

Abstract Handling more than thousand fossil foraminifera data is very challenging, especially for old-way identification. Determining morpho-taxonomy by conventional microscopic observation is very time-consuming and can lead to innacuracy identification. We are certain that the process could be advanced through big data analysis using a machine learning approach. Foraminifera fossils have already become a common standard for biostratigraphic proxies and paleoenvironmental interpretation. Therefore, the objective of this study was to develop an automated identification method using Convolutional Neural Networks (CNN). We used standardized Scanning Electron Microscopy (SEM) images of foraminifera acquired from various open-source databases for image classification. The analysis was conducted using Python programming language in Google Colaboratory. The image dataset is categorized based on its genus (n: 138) and divided into training and test/validation data for accuracy simulation (total image: 1833; data training: 1387; test/validation: 237/237). The best-fit accuracy values of the training-data and testing-data were between 97-86%:73-77%, with parameters including epoch number ranging up to 40, learning rates of 0.05, and a batch size of 64. These values indicate good prospects for foraminifera SEM Image taxonomic classification, demonstrating a noteworthy level of identification accuracy (63%). The outcomes of this study offer a new method for further effective automated morpho-taxonomic identification of foraminifera fossils using other conventional optic microscopy.