Deep-sea imagery is a non-destructive tool for monitoring seafloor litter, but manual inspection limits its scalability. DeepLitterAI, a YOLOv11x-based detector combined with BoT-SORT tracking, was developed to automatically detect and quantify deep-sea benthic macrolitter. The model was trained with J-Litter, a dataset of 12,029 annotated images extracted from footage acquired by the Japan Agency for Marine-Earth Science and Technology (JAMSTEC) between 1983 and 2025. J-Litter includes small litter objects and non-litter images. DeepLitterAI was compared with a model trained only on large, clearly visible litter to evaluate training-data bias. On realistic test images containing small objects, DeepLitterAI achieved average precision (AP), which summarizes precision-recall performance, of 0.79-0.81 and F1 scores, defined as the harmonic mean of precision and recall, of 0.71-0.76 for plastic films, plastic bottles, and beverage cans (mean AP@0.5 = 0.80 and mean F1 = 0.73), with AP values averaging 1.6-fold higher than those of the large-object model. In six independent survey videos acquired at depths of 860-5641 m, using manual inspection as the reference, DeepLitterAI achieved F1 scores of 0.65-0.89, with a mean of 0.77 across categories. AI-derived total litter density averaged 1.1 times the manual estimate. AI analysis was 1.5-3.1 times faster than manual inspection, with a mean speed-up of 2.1-fold. Performance was robust to viewing angle, whereas false-negative rate increased to 25% under the highest synthetic turbidity condition. DeepLitterAI provides a scalable, non-destructive framework for standardized monitoring and mapping of deep-sea macrolitter.
🌊
