Plastic Plants: Evaluating Model and Phenomenon Transferability of Water Hyacinths as a Proxy for Riverine Plastic Pollution Detection

By Giel Hagenbeek

Abstract
Rivers are major pathways for plastic pollution entering the ocean, particularly in tropical regions. Water hyacinths (WHs), invasive aquatic plants that form widely stretching mats on the river surface, have been shown to entangle macroplastics, offering potential as a proxy for detecting plastic pollution via remote sensing. This study evaluates the transferability of this approach from Vietnam’s Saigon River to Thailand’s Chao Phraya River. The transferability of 1) sentinel-2 as water hyacinth estimator, 2) A YOLOv8 deep learning object detection and 3) the phenomenon of hyacinths as plastic aggregator was tested.

Field data were collected using UAVs, bridge-mounted cameras, visual observations, and physical hyacinth sampling over a 62 km river stretch flowing through Bangkok. Sentinel-2 imagery classified hyacinths using a naïve Bayes approach. Two YOLOv8 deep learning models were first tested on hand-annotated data and then applied to detect 1) water hyacinths, 2) free-floating plastics, and 3) entangled plastics in 21,000+ images.

The water hyacinth detection model transferred well (mAP50 = 68%), while plastic detection varied: entangled plastics were reliably detected (mAP50 = 54%), but free-floating plastics underperformed (mAP50 = 23%). Sentinel-2-derived WH coverage corresponded well with bridge imagery. WHs trapped 32% of plastics on average, peaking at 78% upstream. The trapping ratio (trapped plastics/ all plastics) decreased going further downstream, however the concentration of plastics per hyacinths increases. This suggests that not only hyacinths have the capacity to entangle plastics, but also indicates similar flowing dynamics of aggregation. Findings confirm WHs are effective plastic aggregators, supporting scalable monitoring with transferable models, though site-specific tuning remains essential.