Building a Comprehensive Dataset for Training Object Detection Algorithms applied on Plastic Transport Monitoring in Riverine Environments

Plastic monitoring is a challenging task worldwide. Currently, limited plastic measurements are available along the river in coastal areas or in the ocean. Such data from traditional manual monitoring can contribute to describing plastic transport dynamics within river networks, but not extensively in both spatial and time scales. Consequently, it is crucial to advance long-term monitoring within the river corridor in order to properly quantify and characterize  plastic transport and fates.

Recent advances in optical sensing, using commercially available camera systems (e.g. fixed cameras, drones, smartphones) provide huge opportunities in scene monitoring, which has been already successfully integrated in environment-controlled plastic recycling facilities. In this context, image processing techniques can represent a valuable tool, since their use in natural environments introduces a number of difficulties related to light conditions, shadows, and environmental changes (e.g., riparian and submerged vegetation). Therefore, there is a need to build robust methods able to handle such disturbances balancing detection performance with computational cost. 

Considering all these factors, this work utilizes four river plastic datasets (taken from Indonesia, Italy, The Netherlands, and Vietnam) and explores the possibility of tier-based plastic detection, characterization based on different levels of plastic type (from generalized “plastic” to more specific types e.g., plastic, plastic bag, plastic bottle etc.). These datasets represent very different water systems, e.g. urban water systems, natural rivers, tidal rivers, tropical rivers with diverse levels of lighting conditions, water spectra, camera angle, and image resolutions. Different data combinations and augmentation were explored which were used to train base models of YOLOv7 and YOLOv8 (You Only Look Once family of single detectors). Resulting models were compared in terms of transfer learning performance, labor and computational cost.

This work is part of a PRIN funded project, RiverWatch: a citizen-science approach to river pollution monitoring. Preliminary results show that with constant training parameters (batch=16, epoch=100), YOLOv8 performs better than YOLOv7 in river plastic detection. In fact, even though YOLOv7 provides a higher plastic count, this often includes false positives, with generally lower inference scores than YOLOv8. In addition, simple brightness adjustments appear to have a varying effect in improving detection performance depending on plastic types. 

We presented data augmentation methods and techniques in order to improve algorithm detection performance without complicating its network architecture, also in this way the dataset will remain workable with future algorithms. Future work includes the exploration of adding pre-detection localization layers in the test data to enhance local features prior to detection. 

Keywords: river plastic detection, optical remote sensing, YOLO, tier-based plastic characterization, data augmentation