Advancing Plastic Pollution Monitoring Through Enhanced Protocols and Deep Learning: applicability and effectiveness in real-world scenarios (Le Stang, France)

Plastic pollution is pervasive across all environmental compartments, from mountain ranges to abyssal plains. Among these, beaches—and particularly the wrack line—are recognized as critical sites for monitoring plastic pollution. Established programs, such as the French monitoring program (RNS-mP-P), track meso- and large microplastics along beaches. Building on these efforts in the context of the Free LitterAT Interreg project, this study aims to develop a complementary tool to accelerate and expand data acquisition and formatting for monitoring plastic pollution.

A new acquisition protocol was firstly designed. A survey site was selected in Brittany, France (Le Stang), where Cedre has been conducting active monitoring since 2018. Data were collected between January 2023 and July 2024, with seasonal surveys yielding a comprehensive dataset of 2,169 measurements. The study site comprised a 100-meter stretch along the wrack line, examined using quadrats of 20x20 cm, 40x40 cm, and 80x80 cm, spaced at 1-meter intervals. Photos were captured using a dedicated device designed for consistent replication over time and space.

Then, an integrated processing phase evaluated human factor influences and database representativeness to support deep learning solutions. Photos were interpreted and meso- to large microplastics were classified into five categories: Fiber, Film, Foam, Fragment, and Pellet. Three independent users labeled the data, organizing it into training and validation datasets.

Thirdly, a convolutional neural network (U-Net) was employed to analyze the dataset. A tailored training, testing, and validation strategy was established to optimize the use of the unique dataset.

Results were finally benchmarked against the existing RNS-mP-P networks for microplastic monitoring, and recommendations were proposed. For example, the 20x20 cm quadrat setup, spaced every 2–5 meters, emerged as the best compromise for ease and efficiency in the study context.

This proof-of-concept demonstrates the feasibility of integrating advanced methodologies into existing monitoring frameworks. The approach not only enhances data acquisition but also facilitates large-scale implementation through professional and citizen science initiatives.

The findings underscore the potential of combining field monitoring protocols with machine learning to create effective, scalable strategies for environmental plastic pollution monitoring.