MIPT and the Institute of Oceanology of the Russian Academy of Sciences: Garbage Detection System in the Northern Seas

The main articles are:

• Video analytics (terms, applications, technologies)
• Plastic waste

2026: Developing a system to detect plastic and other types of debris in the northern seas

Specialists from the Moscow Institute of Physics and Technology (MIPT) and the Institute of Oceanology of the Russian Academy of Sciences have created a system based on artificial intelligence that can automatically detect floating marine debris and other objects on the sea surface under Arctic conditions from the ship . The development will allow large-scale monitoring of pollution of the oceans. MIPT announced this on January 15, 2026.

𝓲

Фото: MEPhI

Pollution with plastic and other types of garbage has become one of the main threats to the ecosystems of the world's oceans along with climate change. Of particular concern is the Arctic region, where traces of microplastics are found both in the body of marine life and in bottom sediments. Traditional methods of monitoring large debris on the surface of the sea using visual observation require huge human resources and do not provide the necessary coverage of water areas.

MIPT found a solution to this problem. The system developed by scientists is based on two machine learning approaches: classification of images with contrast learning and direct detection of objects. Both methods were tested on a unique dataset collected during a scientific expedition in the Arctic in the fall of 2023.

We processed more than 500,000 photographs of the sea surface taken in the Barents and Kara Seas. Particularly difficult were the difficult conditions of the survey: sea foam, pitching of the vessel and extensive glare from the sun - which greatly make it difficult to detect small objects on the surface of the water and at a shallow depth. The system is capable of identifying four types of objects: marine debris, birds, glare on the water and drops on the camera lens. The development is especially relevant for the Arctic region, where pollution with insoluble anthropogenic debris poses a growing threat to a fragile ecosystem, "said Mikhail Krinitsky, head of the MIPT's machine learning laboratory in earth sciences, one of the authors of the study.

The most effective approach for detecting marine debris was the ResNet50 + MoCo contrast training and the CatBoost classifier. He showed an accuracy of 0.4 on the metric F1-score. For comparison, the popular YOLO algorithm was only able to achieve accuracy of about 0.1 for this task, although it coped better with bird detection (0.73).

{{quote 'The low efficiency of YOLO may be due to the fact that marine debris is often small objects, poorly visible against the background of waves. Also, fortunately, litter is still quite a rare occurrence. The small number of examples to be detected is a classic problem for machine learning models. Our approach with preliminary isolation of image fragments made it possible to better cope with such a feature of statistical learning, "added Olga Belousova, co-author of the work, junior researcher at the MIPT Laboratory of Machine Learning in Earth Sciences. }}

In the future, scientists plan to improve algorithms for working in real time, increasing specificity for floating marine debris and adapt them for use on autonomous monitoring platforms.

The project is being implemented with the support of the Presidential Nature Fund, grant number EKO-25-2-003542