Identifying River Plastic Hotspots from Space

Rivers play a key role in the transport and retention of floating debris, including plastics. Reliable and scalable monitoring of riverine plastic accumulation is essential for identifying hotspots, understanding debris movement, and supporting mitigation strategies. However, conventional in situ monitoring approaches are often labor-intensive, spatially limited, and difficult to deploy consistently across large or remote river systems. This study presents a semi-automated, image-based monitoring framework that integrates satellite remote sensing and machine learning to detect and map riverine plastic accumulation hotspots at a global scale.

The methodology integrates high spatial resolution imagery for precise manual annotation of accumulation areas and multispectral Sentinel-2 data for classification of litter hotspots using Random Forests in Google Earth Engine. The workflow combines the most influential spectral bands with targeted spectral indices, including NDVI, PI, FDI, and SI13, to enhance class separability between plastic, water, and vegetation.

The methodology is evaluated across three highly polluted river systems in Indonesia, Guatemala, and Ghana. These sites represent a wide range of hydrological and environmental conditions, including floating vegetation, canopy shading, and narrow urban channels affected by pixel mixing. Results demonstrate high within-river classification performance, with overall accuracies up to 99.5% on independent sections of the same river, and robust cross-river generalization when spectral indices are incorporated, achieving plastic F1-scores up to 79%.

In addition to image classification, the workflow supports multi-temporal analysis to generate hotspot frequency maps, enabling the identification of persistent plastic accumulation zones linked to river morphology and infrastructure. Feature-importance analysis highlights the relevance of specific spectral bands and indices across different environmental conditions and supports the development of reduced, generalizable models.

To facilitate reproducibility and large-scale application, the methodology is operationalized in an open-access Google Earth Engine application that enables users to apply the trained model to rivers worldwide using Sentinel-2 imagery. The proposed framework contributes to the advancement of environmental monitoring and provides a foundation for future developments toward global, long-term assessment of river plastic dynamics.

 

More information: https://doi.org/10.1016/j.isci.2025.114570