Trash Detection in Surface Waters Using YOLOv8

Aquatic environments are increasingly polluted by various types of trash, such as bottles, plastics, and cans, which accumulate on water surfaces and pose severe threats to marine habitats. The increasing pollution necessitates proactive monitoring of waterways in coastal cities to detect and mitigate the presence of trash before it reaches the ocean. However, effective detection of floating trash remains challenging due to variations in trash types, sizes, and viewing angles. This study aims to develop and evaluate YOLOv8 models for detecting floating trash in surface waters from multiple perspectives to monitor pollution levels and quantify trash. We trained and assessed YOLOv8 architectures for single-class trash detection using the FloW-Img and WaterTrash datasets, capturing both USV and aerial perspectives to ensure comprehensive detection from various angles. The YOLOv8x model achieved the highest mAP50 scores, with 0.923 on the WaterTrash dataset and 0.834 on the FloW-Img dataset. Despite differences in model size, the performance gap was marginal, demonstrating YOLOv8's robustness. These results highlight the potential for further enhancements in detecting tiny and partially submerged trash, as well as distinguishing trash from reflections and waves.