A study on classification and monitoring of marine debris using multi-spectral images and deep learning

Marine environmental issues due to marine debris are worldwide phenomena. According to a press release from the Ministry of Oceans and Fisheries of Korea, marine waste collection from the coastal area increased yearly. In 2020, it collected 1.38 million tons, about 45% more than in 2018. To remove them, they were collected and monitored through field monitoring systems. However, it is very inefficient in terms of time and cost. Therefore, the remote sensing approach can be suited for classifying and investigating marine waste dumped in coastal areas. Previous studies have classified marine waste by combining remote sensing based on RGB images and artificial intelligence. However, actual marine waste is often damaged, or its shape is difficult to recognize through RGB images. This study was conducted to classify various wastes using multi-spectral camera and a convolution neural network (CNN) model. We first trained and tested CNN model using three wastes, such as a brown paper box, an orange-colored buoy, and a blue plastic basket with different spectral characteristics in the land environment. Then, we conducted the classification of marine waste using CNN model and multi-spectral images taken with Uncrewed Aerial Systems (UAS) in the marine environment around Socheongcho-Ocean Research Station (S-ORS). The CNN model were trained using 1,452 seawater and 1,319 clear plastic images around the S-ORS with 128 x 128-pixel size. We calculated precision, recall, f1-score, and accuracy, suggesting that the CNN model could be used to classify various marine wastes in the various ocean environment. Overall, these results can provide useful information for marine waste monitoring.