Real-Time Riverine Plastic Waste Detection Using an Enhanced YOLO Framework

The accumulation of plastic waste in riverine ecosystems poses severe environmental and ecological threats, necessitating real-time monitoring solutions. This research introduces a novel framework for detecting riverine plastic waste using an enhanced YOLOv8 object detection model. The methodology incorporates a meticulously designed dataset preparation pipeline, leveraging XML annotations converted to YOLO format, and a custom training loop optimized for aquatic environments. Key innovations include advanced hyperparameter tuning strategies, integration of per-epoch performance metrics, and detailed visualizations to analyze model behavior. The dataset preparation process ensures robust training by splitting data into train, validation, and test sets, while annotations are normalized for compatibility with YOLOv8. The system was trained on a diverse dataset using an NVIDIA RTX 3060 GPU, achieving a detection accuracy of 94.6 % with significant improvements in precision, recall, and mean average precision (mAP). Results demonstrate the model's robustness across varying environmental conditions, supported by comprehensive reporting mechanisms that include per-class analysis, F1 score evaluations, and precision-recall curves. Additionally, the study provides insights into ground truth versus predictions, enabling detailed performance analysis across classes. This work contributes to the growing field of AI-driven environmental monitoring, offering a scalable and efficient solution for addressing plastic waste pollution in riverine ecosystems while paving the way for future advancements in realtime waste detection technologies.