Species-level microfossil identification for globotruncana genus using hybrid deep learning algorithms from the scratch via a low-cost light microscope imaging

Abstract

Paleontologists generally use a low-cost electro-optical system to classify microfossils. This manual identification is a time-consuming process and it may take about a long time, especially if there are thousands of microfossil samples. In order to solve this problem, we propose a hybrid method based on Convolutional Neural Networks (CNN) and Bidirectional/Long Short-Time Memory (LSTM/BiLSTM) networks for the automatic classification of Globotruncana microfossil species. First, the images of microfossil samples were collected with a low-cost system and labeled by a paleontologist. After preprocessing, the classification is carried out with different combinations of CNN, LSTM, and Bidirectional LSTM (BiLSTM) models from the scratch developed in this paper. Finally, detailed experimental analyses have been made using accuracy, sensitivity, specificity, precision, F-score, and area under curve metrics. In the existing literature, as far as we know, this study is the first investigation work of prediction Globotruncana microfossil species using hybrid deep learning algorithms. Experiments demonstrate that the proposed models have reached the best accuracy with 97.35% and the best AUC score of 0.968 for automatic identification of Globotruncana microfossil species.