Japan Geoscience Union Meeting 2018

Presentation information

[JJ] Oral

B (Biogeosciences) » B-PT Paleontology

[B-PT06] Biotic History

Sun. May 20, 2018 10:45 AM - 12:15 PM 101 (1F International Conference Hall, Makuhari Messe)

convener:Isao Motoyama(Department of Earth and Environmental Sciences, Yamagata University), Takao Ubukata(Division of Geology & Mineralogy, Department of Earth & Planetary Sciences, Kyoto University), Kazuyoshi Moriya(早稲田大学 教育・総合科学学術院 地球科学専修), Chairperson:Motoyama Isao, Ubukata Takao, Moriya Kazuyoshi

10:45 AM - 11:00 AM

[BPT06-06] Automatic image recognition of microfossils using Artificial Intelligence

*Tatsuhiko Yamaguchi1, Kyoko Hagino1, Yousuke Taira2, Yohei Hamada3, Hitoshi Saito2, Jonaotaro Onodera4, Takuya Itaki5, Tatsuhiko Hoshino6,3, Fumio Inagaki3,7,6 (1.The Center for Advanced Marine Core Research, Kochi Univ., 2.NEC Corporation, 3.Kochi Institute for Core Sample Research, JAMSTEC, 4.Research and Development Center for Global change, JAMSTEC, 5.Geological Survey of Japan, Marine Geology Research Group, 6.Research and Development Center for Submarine Resouces, JAMSTEC, 7.Research and Development Center for Ocean Drilling Sciences, JAMSTEC)

Keywords:Machine learning, Taxonomy, Microfossils, Artificial intelligence

Deep learning, using the multilayered neural networks, is a form of a broader family of machine learning methods learning data representations which is a subset of Artificial Intelligence (AI). A combination of the deep learning method and image recognition potentially enables to recognize microfossil taxa in microscopic images automatically. Here we present the application of the AI software to microfossil images. The AI software "RAPID machine learning" (NEC Ltd.) was applied into recognition of six species of Quaternary calcareous nannofossils: Emiliania huxleyi, Gephyrocapsa oceanica, Gephyrocapsa ericsoni, Reticulofenestra haqii, Gephyrocapsa caribbeanica, and Reticulofenestra product. First, we captured 300 optical polarization images each of the six species. Using a training dataset consisting of randomly selected 30 images for each taxon, an auto-recognition model was built by RAPID. Subsequently, the model predicts taxa in a testing dataset of the other 20 images per each taxon to test if the 6 taxa could be recognized correctly . As a result, the percentage of correct classification of G. oceanica and R. haqii were very high as 100% and 95%, respectively. On the other hand, the other 4 species could not be discriminated, showing that 0–35% of the images were correctly classified.
We also would like to introduce the AI applications for discrimination of microfossils other than calcareous nannofossils and to discuss the possibility of automatic recognition.