Deep Learning-Based Floating Debris Detection for Marine Pollution Monitoring

Marine pollution generated by floating debris is dangerous to aquatic ecosystems and can increase risk factors of marine organisms, necessitating robust debris detection systems for monitoring and management. This paper proposes a floating debris identification system based on the you only look once, version 8 (YOLOv8) model, which is trained on 4000 annotated images classified as dense-debris, sparse-debris, single or isolated debris and partial debris portions. The performance of the model is tested under different objects, image qualities and orientations by introducing three different types of noises which are Gaussian, speckle and salt-and-pepper at increasing densities. To increase the adaptability, noise reduction techniques such as Gaussian and moving average filters are used in this study for achieving a higher detection accuracy. On raw images, the YOLOv8 model outperformed current YOLO versions as well as state-of-the-art models as cascade region-based convolutional neural networks (R-CNN), YOLOv5, optimized YOLOv4-tiny model and YOLOv7, with a precision of 82.85%, recall of 91.10% and mean average precision at 0.5 intersection over union (mAP@0.5) of 97%. The performance of the model is decreased under noisy environments, but the use of filters improved the accuracy significantly, obtaining the accuracy of 96.57% and precision of 98.98% in low-density Gaussian noise by applying the Gaussian filter. Compared to previous YOLO versions, the YOLOv8-based approach showed significant increases in precision, recall and mAP, making it a better solution for debris identification. Results demonstrate that the YOLOv8 model's can achieve a higher accuracy and computational efficiency in real-time marine debris identification under complicated and noisy environments. Evaluation results demonstrate the effectiveness of the proposed method for automatic marine pollution monitoring and management applications.