Authors: Wu C, Chen Z, Peng C, An C
Plastic pollution in aquatic environments has become a growing global concern, with floating macroplastics posing serious ecological, economic, and health risks from inland water bodies to remote coastal zones. Despite advances in satellite remote sensing, reliable and scalable detection of floating plastics remains limited by spectral confusion and the lack of optimized indices tailored for this purpose. In this study, we address this gap by developing two novel spectral indices, Index-1 and Index-5, specifically designed to enhance the spectral separability of macroplastics from natural floating materials such as driftwood and aquatic vegetation in Sentinel-2 imagery. These indices were derived from hyperspectral reflectance measurements of water, wood, and plastic samples and selected from six candidate band combinations with the strongest spectral contrast. Integrated into Random Forest classifiers and evaluated using the Sentinel-2-based Marine Debris Archive (MARIDA) dataset, the inclusion of Index-1 improved the F1 score for plastic detection from 0.7952 ± 0.0119 to 0.7987 ± 0.0170, Index-5 to 0.8166 ± 0.0063, and the combined indices to 0.8211 ± 0.0047. Independent validation using the Plastic Litter Projects (PLP) 2021 testing dataset confirmed these improvements, with higher F1 means for models including Index-5 compared to that of RF1. These results underscore the effectiveness and generalizability of the proposed indices across diverse coastal environments. By enabling more accurate and timely identification of plastic accumulation zones, this work supports targeted cleanup efforts and ecological risk assessment, providing a scalable tool that contributes directly to global environmental monitoring and mitigation of plastic pollution.
Keywords: Machine learning; Macroplastic; New index; Random Forest; Sentinel-2;
PubMed: https://pubmed.ncbi.nlm.nih.gov/41406508/
DOI: 10.1016/j.marpolbul.2025.119009