DeepLitterAI: Automated detection and quantification of deep-sea benthic plastic and macrolitter with field validation in waters around Japan.

DeepLitterAI achieves an average precision of 0.79-0.81 and F1 scores of 0.71-0.76 in detecting deep-sea plastic litter, outperforming models trained only on large objects by 1.6 times.

• Why it matters: Monitoring deep-sea benthic litter is essential for understanding marine pollution, but manual inspection is limited in scalability and efficiency.
• What they did: The team developed DeepLitterAI, a YOLOv11x-based detector with BoT-SORT tracking, trained on 12,029 annotated images from the J-Litter dataset, including small litter objects and non-litter images, and validated it in waters around Japan.
• The result: DeepLitterAI provided robust, scalable detection with a mean F1 score of 0.77 across six survey videos, increased speed by 2.1-fold over manual inspection, and offers a non-destructive method for standardized deep-sea litter monitoring.