Micropaleontology is not only about studying the organisms themselves, rather understanding Earth's past environments, with applications ranging from biostratigraphy to paleoceanography as well as being able to study evolutionary changes within morphospecies in time and space. This field is facing numerous challenges, since the analysis of microfossils demands significant human effort and taxonomic expertise, often leading to inconsistencies in interpretations. This work focuses on the application of using Artificial Intelligence (AI), such as Artificial Neural Networks (ANNs), for automatic image recognition of tropical Atlantic middle Eocene radiolarians. Large datasets have been constructed, in order to train different neural networks and our results show that the neural networks can automatically classify several different classes of radiolarians down to a species level, as well as in many cases being able to identify closely related species and even evolutionary transition morphotypes. It has also been able to correctly identify less broken or blurry radiolarians. It was also successfully applied to automatic image recognition for a biostratigraphic work, which in general could detect more general ages or highly precise bio events. This work includes the use of the classical neural network approaches for analysing visual context such as Convolutional Neural Networks (CNNs) but also includes the use of Spiking Neural Networks (SNNs), which is not as commonly used for automatic image recognition, as CNNs. SNNs resulted in almost or equal amount of accuracy obtained as for CNNs, just that the use is more computational efficient and takes up less memory. There have also been some comparisons using traditional morphometric analyses, such as Linear Discrimination Analysis (LDA), giving approximately the same kind of results. Our research not only aims to simplify and speed up the analysis process but also helps in increasing the accuracy and consistency of micropaleontological interpretations, which eventually, will contribute to the high-resolution studies in order to understand Earth's past history.
M. Taniel DANELIAN Université de Lille Directeur de thèse, M. Fabrice CORDEY Université Claude Bernard Lyon1 Rapporteur, Mme Rie HORI Ehime University Examinatrice, M. Thibault DE GARIDEL-THORON Aix-Marseille Université Rapporteur, M. Pierre BOULET Université de Lille Co-directeur de thèse, Mme Allison HSIANG Stockholm University Examinatrice, Mme Catherine CRONIER Université de Lille Examinatrice, M. François DANNEVILLE Université de Lille Examinateur, M. Philippe DEVIENNE Université de Lille Invité.
français
English