Satellite-based marine pollution monitoring for African coastal energy infrastructure: Transformer semantic segmentation of Sentinel-2 multispectral imagery on the MADOS benchmark
This study demonstrates that general-purpose transformer models, specifically SegFormer-B2 and Mask2Former adapted for four-channel Sentinel-2 imagery, outperform the purpose-built MariNeXt baseline in detecting oil spills and marine debris within the MADOS benchmark, offering a robust and transferable solution for monitoring pollution around African coastal energy infrastructure despite limitations in identifying sparse features due to reduced spectral bands.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The waters off the coast of Africa, particularly along the Atlantic seaboard where oil wells, pipelines, and shipping terminals cluster, face a persistent threat from chronic pollution. For decades, monitoring these waters has been difficult. Field teams struggle to reach remote areas, and thick cloud cover often blocks the view from satellites, leaving gaps in our knowledge of where oil spills or floating debris might be hiding. To fill these gaps, scientists rely on a fleet of European satellites known as Sentinel-2. These satellites circle the Earth every five days, capturing detailed images of the surface using thirteen different colors of light, ranging from the visible spectrum we can see to invisible infrared bands. By analyzing how different materials reflect these specific colors, researchers can distinguish between clean water, oil slicks, plastic waste, and natural floating matter. The challenge has always been turning these raw images into reliable maps that automatically identify pollution, a task that usually requires building complex, custom-made computer programs designed specifically for this one job.
A new study asks a simpler, more practical question: do we really need to build custom programs from scratch, or can we use powerful, general-purpose computer models that have already learned to see the world? These general models, known as transformers, were originally trained on millions of everyday photographs to recognize cats, cars, and landscapes. The researchers wanted to see if these adaptable tools could be repurposed to spot marine pollution just as well as, or better than, the specialized systems currently in use. They focused on a specific dataset called MADOS, which contains hundreds of satellite images of the ocean annotated with over a million labeled pixels showing everything from oil spills and ships to jellyfish and foam. The goal was to test whether a general model could learn to spot the most dangerous pollutants—oil and debris—without needing the expensive, custom engineering that usually accompanies such tasks.
The researchers took two of the most popular general-purpose models, SegFormer and Mask2Former, and made a small but crucial adjustment to let them look at the satellite data. Since the standard models expect three colors (red, green, and blue) like a normal photograph, but the satellite data included a fourth invisible color called near-infrared, the team had to teach the models how to process this extra channel. They did this by carefully expanding the models' initial "eyes" to accept four inputs instead of three, using the knowledge the models already had about visible light to help them understand the new infrared data. They then trained these adapted models on the official MADOS dataset, letting them learn to distinguish between fifteen different types of ocean features, including oil spills, marine debris, ships, and various water conditions.
The results were striking. One of the general models, SegFormer, outperformed the best-known custom-built system designed specifically for this task. While the specialized system managed to correctly identify about 64 percent of the different pollution types and water features on average, the adapted general model achieved a score of 73 percent. More importantly, when it came to the specific pollutants that matter most for protecting coastal energy infrastructure, the general model was significantly more accurate. It identified oil spills with a success rate of 94 percent and marine debris at 81 percent, whereas the specialized system struggled with debris, getting it right only 51 percent of the time. The other general model tested, Mask2Former, also detected oil spills very well, at 95 percent accuracy, though it was less successful at identifying other features overall.
However, the study also revealed where these general models hit a wall. They performed poorly on certain types of floating matter, such as sparse floating algae, foam, and jellyfish. The researchers traced this failure to the specific colors of light they chose to feed the models. By using only four bands of light to keep the system simple and fast, they missed out on the specific infrared colors that are best at distinguishing these delicate, biological materials from the surrounding water. The specialized system, which had access to a wider range of light colors, was better at spotting these tricky items. This suggests that while general models are incredibly powerful, they still need the right "eyes" to see every type of pollution.
The implications for monitoring African coastal waters are significant. The study demonstrates that agencies do not necessarily need to invest in building unique, custom software to track oil spills and debris. Instead, they can rely on freely available satellite images and robust, open-source models that are easier to maintain and update. For regions like the Gulf of Guinea, where cloud cover often obscures the view, this approach offers a practical path forward. By pairing these optical detectors with radar systems that can see through clouds, authorities could create a reliable, continuous monitoring network. The research confirms that a well-trained, general-purpose tool can serve as a strong, cost-effective foundation for protecting vulnerable coastal ecosystems and energy infrastructure, provided the system is tuned to look at the right parts of the light spectrum.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.