Real-time detection of macroplastic pollution in inland waters: development of a lightweight image recognition system

Abstract

Plastic pollution in freshwater ecosystems poses a growing environmental threat, yet the availability of efficient and scalable monitoring solutions remains limited. This study presents a lightweight, real-time macroplastic detection framework based on the YOLOv8 object detection model, optimized for continuous monitoring using video footage from fixed (pontoon-, bank-, or bridge-mounted) camera systems or mobile (Unmanned Aerial Vehicle, UAV-based) deployments for pollution assessment. The model’s performance was evaluated across multiple environmental scenarios, including simulated pollution and real-world UAV footage under moderate and high plastic pollutant loads. To address key challenges such as small object size and occlusion by vegetation, pre-processing techniques including image tiling and blurring were applied. These enhancements led to notable improvements in recall and mean Average Precision (mAP) scores. The proposed system architecture supports both decentralized (on-site) and centralized processing configurations, allowing flexible deployment across diverse monitoring contexts. Beyond its operational applicability, the system enables the large-scale collection of pre-annotated datasets, supporting future model refinement and site-specific training. When combined with hydrological and meteorological data, the resulting time series may serve as a foundation for predictive models of plastic pollution transport, offering a valuable tool for mitigation efforts and early warning systems.