Scientists from the Federal University "Immanuel Kant" have developed an optimal method for training a neural network to automatically recognize floating debris in the sea.
The Russian Ministry of Education and Science explained that this is done through video recordings taken from aboard research vessels.
A ministry spokesperson said that scientists from Immanuel Kant Baltic Federal University, in collaboration with researchers from the Moscow Institute of Physics and Technology and Moscow State University, have developed a method to train a neural network to detect floating plastic waste in the sea. The researchers presented a model capable of distinguishing between plastic, birds, light reflections, and water droplets on a camera lens, enabling its use in continuous monitoring of marine conditions.
According to experts, the global ocean receives up to 23 million tons of human-generated waste annually, causing serious damage to ecosystems.
The researchers said that marine organisms may ingest large plastic particles that obstruct breathing, and may also become entangled in plastic bags, nets, and fibers, making it essential to track and remove marine debris in a timely manner.
This waste is usually detected by scanning the ocean surface from ships, but this process is time-consuming and requires significant human effort, making monitoring large marine areas a complex task.
Therefore, scientists have proposed an alternative based on analyzing images of the sea surface taken by drones or cameras on board ships, using neural networks capable of distinguishing between plastic waste and other elements such as marine animals, foam, and light reflections.
To train and test the algorithms, scientists used video footage collected aboard the research vessel Dalniy Zelentsy during a 2023 polar expedition. This footage, totaling 136 hours, was divided into separate frames, resulting in over half a million images of the sea surface. Of these, approximately 10,000 were manually categorized to identify birds, debris, reflections, and water droplets on the lens.
The algorithm was then trained on pre-tagged data, enabling it to recognize the same objects in untagged images. The researchers also deliberately altered the ratio of empty frames to those containing different objects to improve the model's generalization ability.
The results showed that the neural network that underwent this type of "self-learning" was 30% more efficient at detecting waste compared to models that were trained only on manually labeled data.
