[M-IS20] 地球科学としての海洋プラスチック

[MIS20-05] Quantifying the degree of pollution of beaches using an AI technique

*日髙 弥子¹、杉山 大祐¹、松岡 大祐¹、荒木 文明¹、蒲地 政文¹、石川 洋一¹ (1.国立研究開発法人　海洋研究開発機構)

キーワード：人工漂着物、海岸、ゴミのモニタリング、人工知能、深層学習、物体検出

Beached marine litter is harmful to both humans and nature. Monitoring litter on beaches solely through human effort is limiting in terms of both time and expenditure constraints. Establishing technologies that can quantify beached marine litter automatically will contribute to mitigation strategies and scientific understanding of litter. Therefore, we have attempted to develop an AI (Artificial Intelligence) system that can automatically estimate the degree of pollution of beaches using a deep-learning method. The core technology is “semantic segmentation”, which gives pixel-wise classification on images. For the training dataset, pairs of the original image and corresponding labels (ground truth) were prepared. We received monitoring data from the local government (Yamagata prefecture) and prepared the training dataset for the deep-learning model based on these images. For the labels, each region of an image is categorised into eight classes: sky, sea, sand beach, artificial litter, natural litter, installed object, and natural object. Using this dataset, we trained a deep-learning model that can classify an image into the eight classes. The mean IoU (Intersection over Union) between predicted results and ground truth was 76% for all classes. We will present preliminary results from the model, and would also like to discuss the detection of plastic litter.