[M-IS20] 地球科学としての海洋プラスチック

[MIS20-06] 市民科学によるビッグデータと深層学習を活用した街中ごみ定量化手法の構築

*加古 真一郎¹、菅野 晋太郎²、磯辺 篤彦³、九鬼 正人⁴、山本 虎太郎⁴、小嶌 不二夫⁴ (1.鹿児島大学大学院理工学研究科、2.鹿児島大学工学部、3.九州大学応用力学研究所、4.株式会社/一般社団法人 ピリカ)

キーワード：市民科学、深層学習、スマホアプリ ピリカ、ビッグデータ、街中ごみ

In this study, we constructed a deep learning model using training data consisting of litter images in cities collected by citizen science using a smartphone application “Pirika.” Our deep learning model can classify and quantify the amount of litter in cities into various categories (plastic bags, plastic bottles, cans, and others). The model’s accuracies of object detection and classification were 83.0% and 82.9%, respectively. However, these accuracies of the model are guaranteed only under limited conditions and must be improved for practical use. The results of the learning curve shows that there is a limit to learning with the current training data and raises a concern that the model overfits the training data (validation accuracy of 70% and training accuracy of 99%). In future studies, the versatility of the model must be improved by increasing the size of the learning dataset by including various patterns for detection and classification. The smartphone app “Pirika” must also be updated with a function that disallows taking images that might lower the accuracy of our model.