Computer Vision / Video Analytics

High-Tech AI Framework Transforms Global Marine Pollution Tracking

A satellite in space with earth in the background.

An AI-powered remote sensing study offers a dynamic new tool for global ocean cleanup efforts. Detailed in the ISPRS Journal of Photogrammetry and Remote Sensing, the breakthrough unveils MariNeXt, a deep-learning framework that detects and identifies marine pollution using high-resolution Sentinel-2 imagery. MariNeXt could revolutionize how resource managers and agencies globally monitor and mitigate marine pollution by accurately detecting marine debris and oil spills on the sea surface.

“Marine debris is currently considered one of the most pressing issues in marine pollution. The ability to automatically and accurately identify debris is important for responding to and reducing the significant threats to ecosystem health and the blue economy,” said Katerina Kikaki, study lead author and postdoctoral researcher at the National Technical University of Athens. 

Sources of pollution such as oil spills, marine litter, and algal blooms are an ongoing threat to human health, aquatic life, and the economy. In the past, detecting marine pollutants using manual methods was labor-intensive and time-consuming, resulting in only a fraction being identified. 

“AI is an increasingly powerful tool for ocean monitoring. Combined with remote sensing, it offers automated data collection and analysis across large spatial and temporal scales, enabling more comprehensive and cost-efficient monitoring,” Kikaki said.  

Effective marine pollution monitoring systems are crucial for achieving UN Sustainable Development Goals, as they play a key role in ensuring the long-term health of marine environments. However, current AI algorithms fail to identify pollutants accurately. 

Most proposed methods have been designed to detect a single marine pollutant or a small number of sea surface features. Plus, they tend to operate locally, without the ability for large-scale monitoring. Another challenge is that marine pollutants have complex optical properties, and current satellite sensors aren’t always equipped to handle them.

Looking ‌to overcome these limitations, researchers from the National Technical University of Athens and King Abdullah University of Science and Technology developed MariNeXt. The deep learning framework integrates advanced data augmentation techniques and a multi-scale convolutional attention network, enabling it to learn and generalize from wide-ranging conditions and sea surface features.

The researchers trained MariNeXt on the Marine Debris and Oil Spill (MADOS) dataset, which they also created using about 1.5M annotated pixels from 174 satellite scenes collected worldwide between 2015 and 2022. The comprehensive dataset includes 15 classes, including floating marine debris, oil spills, Sargassum macroalgae, natural organic material, ships, sea snot, and water-related conditions such as waves and turbid or shallow water.

A grid showing satellite image patches of the 15 classes.
Figure 1. An overview of the MADOS patches showing marine pollutants and sea surface features annotated under various weather and sea state conditions

The researchers developed and tested the model using the cuDNN-accelerated PyTorch framework on two NVIDIA RTX A5000 GPUs, each with 24 GB of memory. The researchers were awarded the two RTX A5000 as recipients of the NVIDIA Academic Hardware Grant Program.

According to study coauthor Ioannis Kakogeorgiou, “The significant GPU capacity enabled the team to develop advanced deep-learning solutions beyond traditional machine-learning methods like random forests. This high-performance hardware allowed extensive experimentation with larger models, higher input resolutions, and increased batch sizes.”

The MariNeXt model reached an overall accuracy of 89.1% in identifying marine pollutants and sea surface features across different ocean conditions. The AI framework also produced promising predictive maps and outperformed other machine learning baseline models, highlighting its potential for understanding and monitoring oceanic environments.

While MariNeXt is a useful tool for ocean monitoring, it has limitations. The dataset is by nature unbalanced and some classes like marine water and oil spills, are abundant, while others, like foam and natural organic material, are less represented. 

This could limit the model’s ability to detect ‌less-represented classes in regions beyond the dataset’s coverage. The researchers are currently working on improving MariNeXt’s predictive capabilities.

“Putting the limitations aside, MADOS is a valuable dataset that benchmarks machine learning algorithms for marine pollution detection from open Sentinel-2 data, supporting the development of future operational marine monitoring solutions,” Kikaki said. 

Learn more and access the open-source code on GitHub.

Read the research Detecting Marine Pollutants and Sea Surface Features with Deep Learning in Sentinel-2 imagery.

Discuss (0)

Tags