Satellite-Based Deep Learning for Marine Plastic Detection and Drift Prediction

Abstract

Marine plastic pollution presents a growing environmental challenge, particularly in coastal regions where debris accumulation threatens marine ecosystems and biodiversity. This paper proposes a satellite-based deep learning framework for detecting floating plastic debris and predicting its movement across ocean surfaces. The proposed method integrates multispectral satellite imagery, ocean current information, and meteorological observations through a spatiotemporal neural network. The experimental design incorporates 26,000 annotated satellite image patches collected from 14 coastal regions, together with trajectory information from 860 ocean monitoring drifters. In the illustrative evaluation, the proposed detection model achieves a mean Intersection over Union of 0.79, compared with 0.68 for a conventional segmentation baseline. For 72-hour debris movement prediction, the framework reduces average displacement error from 15.8 to 9.4 kilometers. The system also achieves a detection recall of 0.88 under moderate cloud coverage and variable sea-surface conditions. These findings suggest that integrating remote sensing with spatiotemporal learning can improve the identification and movement prediction of floating marine debris, supporting more targeted ocean pollution monitoring efforts.