[M-IS17] Ocean Plastics, an earth science perspective

[MIS17-01] Dataset for beach plastic litter detection using AI and the preliminary results from the use

*Mitsuko Hidaka^1,2, Koshiro Murakami¹, Shintaro Kawahara¹, Daisuke Sugiyama¹, Shinichiro Kako^2,1, Daisuke Matsuoka^1,2 (1.Japan Agency for Marine-Earth Science and Technology , 2.Kagoshima University)

Keywords:Beach monitoring, Macroplastic, Image processing, Artificial intelligence, Deep learning, Training dataset

Marine plastic litter pollution is a global concern. However, the technologies to accurately estimate pollution's seriousness have rarely been established yet in any environment. Beaches are places where human living areas, and a variety of litter are washed ashore from the river and offshore. To estimate the standing stook or load late of the pollutant on the beaches, relatively laborious approaches have been taken for a long time, such as a manual collection of the litter or visual transect by human effort. Applications of images taken by a camera installed on unmanned aerial systems, webcams, and smartphone cameras have the possibilities to provide alternative monitoring methods to such labour works, and automated analysis methods for the images are essential. We have already established a method to classify artificial litter on beaches at pixel level using a deep learning technique, and as the next step, attempting to develop a technology to classify macro plastic objects in 13 classes. To develop the technology using deep learning, preparation for the training dataset is vital because it directly affects the model performance. We produced a beach plastic litter training dataset for machine learning use, namely the BePLi Dataset (Beach Plastic Litter Dataset) and version 1 is already published in Sea Open Scientific Data Publication (SEANOE). Moreover, version 2 is now being prepared. The dataset utilised a monitoring record provided by one of the local Japanese governments (Yamagata Prefecture), which contains pictures of the beach images from 167 stations. These were all taken by humans standing on the beach and were taken from different backgrounds, such as sand beaches, rocky beaches, and tetrapods. The image records were pasted on Excel monitoring record files, and for the BePLi Dataset, 3708 images were extracted from the Excel files. The corresponding annotations of instance segmentation to the images were provided by manual annotation. These datasets can be used for the development of instance segmentation and bounding-box-based object detection. The label composition of the BePLi Dataset was analysed, and deep learning models were trained using the dataset. We will introduce the details of the dataset and preliminary results of the deep-learning model using the training dataset.