AI Beneath the Surface: Japanese Researchers Develop DeepLitterAI to Track Seabed Plastic Waste

YOKOSUKA, JAPAN — In a major technological leap for ocean conservation, scientists at the Japan Agency for Marine-Earth Science and Technology (JAMSTEC) have developed an artificial intelligence system designed to automatically detect, classify, and count plastic waste scattered across the deep ocean floor.

Dubbed DeepLitterAI, the model tackles one of marine biology’s most notorious bottlenecks: manually scrubbing through thousands of hours of subsea submersible footage to monitor marine litter.

Man-made debris on the ocean floor. Source: Marine Conservation Institute

Overcoming the “Needle in a Haystack” Challenge

While floating ocean plastic garners widespread public attention, an estimated majority of marine plastic debris eventually sinks into the abyss. Deep-sea trenches and ocean floors act as terminal accumulation sinks, yet cataloging this debris has historically relied on laborious, frame-by-frame visual inspections by researchers.

Existing computer vision models struggled in deep-sea conditions. Deep-sea submersibles—such as JAMSTEC’s SHINKAI 6500 and remotely operated vehicles like HYPER-DOLPHIN—use wide-angle lenses to navigate pitch-black waters. Under these lenses:

• Debris often occupies just 5% to 10% of the frame width.
• Man-made plastics easily blend in with rocks, sediment, corals, and bottom-dwelling marine organisms.
• Varying light falloff and water turbidity frequently trigger false positives.

Trained on Four Decades of Seabed Surveys: The J-Litter Dataset

To build a model tailored specifically for real-world ocean exploration, the research team—led by JAMSTEC biological oceanographer Ryota Nakajima—compiled a new visual benchmark named J-Litter.

Drawing from deep-sea footage collected across Japan’s surrounding waters since 1983, the team curated and annotated more than 12,000 images. Crucially, the training data integrated:

1. Low-resolution and distant fragments of plastic bags, containers, and synthetic ropes.
2. Natural “distractors,” including irregular rock formations, sponges, and deep-sea fauna that previous models routinely mistook for garbage.
3. Algorithmic augmentations like inversion, motion blurring, and light distortion to prepare the AI for volatile underwater visibility.

Slashing Analysis Time from Months to Days

Published in the environmental science journal Environmental Pollution, the study’s findings demonstrate that DeepLitterAI achieves human-level detection benchmarks at commercial processing speeds:

“Because huge quantities of plastic sink directly to the seafloor, assessing the true scale of pollution has always been exceptionally difficult,” said lead researcher Ryota Nakajima. “With this tool, we can rapidly identify accumulation hotspots and provide the hard empirical data needed to target conservation measures.”

A Blueprint for Global Marine Monitoring

As international negotiators work toward legally binding global treaties on plastic pollution, scalable monitoring tools are critical. JAMSTEC plans to use DeepLitterAI to automate the processing of archival deep-sea footage worldwide, paving the way for comprehensive global maps of seafloor waste hotspots.