Japanese researchers develop AI to detect plastic waste on the seabed
A team of Japanese researchers has developed an artificial intelligence (AI) system to better detect marine plastic waste on the seabed, as reported in an international journal on Thursday.
The team from the Japan Agency for Marine-Earth Science and Technology (JAMSTEC) developed the system, named DeepLitterAI.
The system is capable of processing data at nearly twice the speed of humans and is expected to be used for real-time monitoring of marine plastic waste.
The team compiled a dataset consisting of approximately 12,000 images based on Japanese seabed recordings taken by the agency since 1983.
This dataset includes small-sized waste objects as well as non-waste objects such as rocks and living creatures, which are often misidentified.
The team then trained the system using various patterns, such as blurring effects and image flipping, to reduce detection errors.
Although most plastic waste entering the ocean eventually sinks to the seabed, it is difficult to obtain an accurate picture of its condition.
“We can quickly identify areas where waste is accumulating in large quantities and use this information to implement mitigation measures,” said Ryota Nakajima, a biological oceanographer on the team.
In tests using original footage, the system was able to identify the type and amount of waste even when objects appeared very small, covering only 5 to 10 per cent of the image width, and successfully detected 80 per cent of major waste types, such as plastic bottles and polyethylene bags.
The error rate compared to visual inspection by experts was approximately 10 per cent, and the system was able to complete the analysis in just a few days, a task that usually takes about one month if performed by humans.
Given that deep-sea recordings are taken using wide-angle lenses, previous AI systems were only able to recognise large-scale waste, often missing smaller debris.
The research team reported that this new method has improved accuracy by 1.6 times.