A comparative study of deep learning-based network model and conventional method to assess beach debris standing-stock.

The conventional survey of marine debris standing-stock has various drawbacks such as high cost and inaccuracy because the total amount of debris in the whole beach is inferred using the results of the manual investigation in selected narrow areas. To overcome the disadvantages, an automatic detection method using a deep learning-based network model was developed to detect and quantify the beach debris. The network model developed in this study classified items with a precision of 0.87 (87%) mAP and showed <5% error compared to actual survey. This study is the first fieldwork in Korea that shows the difference between automatic and conventional methods to predict the beach debris standing-stock. The results provide essential information for the development of effective beach debris management systems and policies.