Computer Vision-based Marine Debris Detection Using Convolutional Neural Network

Abstract

Marine debris has become a serious and growing threat to marine ecosystems, human health, and maritime activities and economies. Various manual monitoring efforts that has been carried out so far are often limited in terms of spatial coverage, efficiency, and resource effectiveness. With the rapid advancement of remote sensing technology and the integration of artificial intelligence, marine debris monitoring can now be automated through computer vision approaches. This study develops a computer vision-based system for marine debris detection using Sentinel-2 satellite imagery and Convolutional Neural Network (CNN) to adhere to the blue economy framework. The proposed approach applies semantic segmentation using the U-Net architecture. The primary dataset used in this study is provides multispectral imagery with a spatial resolution of 10 meters and annotations for four main classes: marine debris, organic material, seawater, and other objects. The imagery enhanced using spectral indices such as the Floating Debris Index (FDI) and the Normalized Difference Vegetation Index (NDVI) helps distinguish spectral characteristics between debris and non-debris classes more clearly. Model performance is evaluated using metrics including accuracy, precision, recall, Intersection-over-Union (IoU), and F1-Score. The best model achieved scores of 0.95, 0.81, 0.67, and 0.77 for each respective metric, demonstrating U-Net's effectiveness in detecting marine debris. The final system is deployed through an interactive Streamlit interface, allowing users to upload satellite imagery, view segmentation results, visualize spectral indices, and preview bounding boxes that highlight detected debris locations. This approach is expected to serve as an effective and adaptive tool to support sustainable marine environmental policies and decision-making.