- 1University of Naples Federico II, Department of Civil Engineering and Environmental (DICEA), Napoli, Italy (khimcathleen.saddi@unina.it)
- 2Scuola Universitaria Superiore IUSS Pavia, Pavia, Italy;
- 3Department of Civil Engineering and Architecture, Ateneo de Naga University, Naga, Philippines;
- 4Dipartimento di Informatica - Scienza e Ingegneria, Università di Bologna, Italy;
- 5Department of Geography, University of Zurich, Zurich, Switzerland;
- 6Hydrology and Environmental Hydraulics Group, Wageningen University, Wageningen, The Netherlands
Recent advances in hydrological monitoring using different camera systems provide a huge potential in long-term monitoring of plastic transport, which is necessary to find the plastic sources and to monitor any progress in efforts to reduce riverine plastic transport. The high interest in using machine learning in different environmental monitoring applications allowed the fast development of models aimed to translate manual visual to computer vision monitoring. However, there is still a lack of robust plastic image datasets that could support machine learning models to detect different plastic classes (i.e., plastic bag, plastic bottle, plastic straw, etc.) that are found in the environment.
In this study, we aimed to identify which data features could be useful to enhance the capabilities of the YOLO series of models (i.e., YOLO World, YOLO NAS, YOLOv8, YOLOv10, YOLOv11) initially trained using a merged dataset (999 images, 15,212 annotations, and 13 plastic classes) taken from different countries (Indonesia, The Netherlands and Vietnam). In addition, we used crowd-sourced images data of river plastics collected with the CrowdWater app (https://crowdwater.ch/), a citizen science app that allows users to report plastic pollution in water bodies. The data was fed to the models for detection 0 (first plastic detection which generates initial labels for iterative training later), in which those learned are considered redundant and unlearned essential–auto image curation. These labels were validated through manual label curation and adjustment. The essential data was added to the existing dataset to fine tune the set of models and the auto image curation will be run again for at least 10 iterations. The performance of these models has been compared for the base dataset (existing and all crowd data) and the optimized dataset (existing and curated crowd data).
This work leverages the value of utilising crowd-sourced diverse data, without the need for a big dataset or a complex algorithm architecture, to implement river plastic detection from local to global scale in the future.
Keywords: river plastic monitoring, crowdwater, image-based object detection
How to cite: Saddi, K. C., Miglino, D., Moe, A. C., Caramiello, C., Poggi, M., van Meerveld, I., van Emmerik, T. H. M., and Manfreda, S.: The value of Crowd-sourced data in Image-based River Plastic Detection, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18045, https://doi.org/10.5194/egusphere-egu25-18045, 2025.
