River plastic hotspot detection from space

TL;DR

This paper introduces a satellite-based method to detect plastic pollution hotspots in rivers using machine learning and cloud computing.

Contribution

A novel cloud-based pipeline using satellite data and machine learning to detect river plastic hotspots with high accuracy.

Findings

The method achieves up to 99.5% accuracy within rivers and 79% F1-score across rivers.

The approach is tested in three diverse river systems with promising results.

An open-access application is released for global plastic monitoring.

Abstract

Plastic pollution threatens terrestrial and aquatic ecosystems, and rivers play a central role in transporting and retaining plastics across landscapes. Effective mitigation requires scalable methods to identify riverine plastic accumulation hotspots. Here, we present a semi-automated, cloud-based pipeline that integrates satellite remote sensing and machine learning to detect river plastic hotspots. High-resolution PlanetScope imagery is used to annotate training regions, which are transferred to Sentinel-2 multispectral data to train Random Forest classifiers within Google Earth Engine. The approach is evaluated across three contrasting river systems—the Citarum (Indonesia), Motagua (Guatemala), and Odaw (Ghana)—to assess transferability under diverse environmental conditions. Intra-river transfer achieves up to 99.5% accuracy, while optimized inter-river transfer yields a plastic F1-score of 79%, outperforming previously reported results of 69%. By providing an open-access Google Earth Engine application, this work enables reproducible, large-scale monitoring of riverine plastic pollution and supports the development of global, satellite-based assessment strategies.

•Developed a GEE-based workflow for river plastic detection using Sentinel-2 data•Used PlanetScope imagery to generate training ROIs for classifier development•Achieved up to 99.5% intra-river accuracy and 79% F1-score in inter-river transfer•Released an open-access application for global Sentinel-2 based plastic monitoring

Developed a GEE-based workflow for river plastic detection using Sentinel-2 data

Used PlanetScope imagery to generate training ROIs for classifier development

Achieved up to 99.5% intra-river accuracy and 79% F1-score in inter-river transfer

Released an open-access application for global Sentinel-2 based plastic monitoring

Earth sciences; Environmental science; Pollution; Remote sensing

Figures