Japan Geoscience Union Meeting 2022

Presentation information

[J] Oral

M (Multidisciplinary and Interdisciplinary) » M-IS Intersection

[M-IS19] Ocean Plastics, an earth science perspective

Sun. May 22, 2022 9:00 AM - 10:30 AM 106 (International Conference Hall, Makuhari Messe)

convener:Atsuhiko Isobe(Kyushu University), convener:Kiichiro Kawamura(Yamaguchi University), Yusuke Okazaki(Department of Earth and Planetary Sciences, Graduate School of Science, Kyushu University), convener:Masashi Tsuchiya(Research Institute for Global Change, Japan Agency for Marine-Earth Science and Technology), Chairperson:Atsuhiko Isobe(Kyushu University), Kiichiro Kawamura(Yamaguchi University)

9:25 AM - 9:40 AM

[MIS19-02] Application of deep learning to the macro beach litter quantification

*Mitsuko Hidaka1, Daisuke Sugiyama1, Koshiro Murakami1, Shin'ichiro Kako2, Daisuke Matsuoka1 (1.Japan Agency for Marine-Earth Science and Technology , 2.Kagoshima University)

Keywords:Beached litter, Macroplastic, Artificial intelligence, Deep learning, Volume estimation, Beach monitoring

Beaches are the place where the human living areas and the ocean environments meet and a variety of litter are washed ashore from the river and offshore. Plastic litter comprises 70-90% of beached litter, and they are thought to be fractured into microplastics by the effect of ultraviolet, winds, and waves. Therefore, understanding the flux of the macroplastic litter on the beach has significant meaning to estimate the total amount of the emission of microplastics into the ocean from the beach. As the first step to establishing the method to estimate the total amount of macro plastic litter on the beach, we have attempted to establish a method to estimate the total amount of artificial macro litter using a deep learning approach. We prepared a training dataset for the machine learning based on 3500 beach images, and we developed a deep learning model which gives pixel-level image classification to the beach images into 8 classes which are including artificial litter and natural litter classes. The accuracy of the model was evaluated qualitatively and quantitatively. The detection accuracy of the model for artificial litter on the test data was around 80%, and the usefulness of the method was demonstrated by the results on the images taken in different areas from training images were taken. Moreover, we showed the potency of the method for cover area estimation by comparing the results by projection transform from the ground image inference and drone image ground truth results. The training dataset that was used in this research is now public from SEANOE - Sea Open Scientific Data Publication. We will present our latest research as above and the outlook for the plastic litter estimation.