A German-Israeli research team used Copernicus Sentinel-1 data to train a deep-learning based oil spill detection system in the South-eastern Mediterranean Sea, which can be used for early-stage oil contamination alerts.
As maritime traffic increases in the Mediterranean Sea, so too does the threat of oil pollution to the region’s marine environment. Its effects on marine mammals and birds are devastating, with inhalation of volatile petroleum causing respiratory irritation and narcosis.
While much pollution is caused by oil spills from tanker accidents, the majority of human-caused oil pollution is in the form of illegal discharges, such as oily ballast water, tanker washing residue and fuel oil sludge.
Early oil spill detection is not only useful to protect marine life, but also to track illegal and deliberate oil pollution in such heavily marine trafficked regions.
Researchers at the SAR Oceanography Team at the Remote Sensing Technology Institute, German Aerospace Center (DLR), aim to assist early-stage oil contamination surveillance with their detection system.
The team developed a deep learning based oil spill early warning system using Synthetic Aperture Radar (SAR) images collected from the Sentinel-1 mission of the European Union’s Copernicus Programme [1].
The researchers chose the South-eastern Mediterranean Sea as their study area since it is a well-known oil spill hotspot.
The region is an important oil transit centre and provides the shortest shipping route from Asia to Europe. Industrial oil and gas activities have escalated here since the 2010s, when large gas fields were discovered.
Oil spills can be detected by SAR
"We chose to use Sentinel-1 imagery not only because they were freely accessible, but also because a large dataset is required to train the deep learning architecture. Only with such a large dataset could we detect not only the large oil spill incidents, but also deliberate oil spills” says researcher Yi-Jie Yang, from DLR, Germany.
The Copernicus Sentinel-1 images were first pre-processed and cropped into smaller images and - if necessary - downscaled to fit the image input size of the object detector.
Subsequently, the oil objects were categorised into size groups (large, medium or small) to enable the detector to perform extra data analysis on specific groups. Objects below a certain size were disregarded in training the architecture.
“The challenges in detecting oil spills are that non-nominal variants can be confused with look-alikes in the radar images, such as low wind areas or wave fronts. Studies applying conventional algorithms, which detect all dark formations first and then classify the oil spills and look-alikes, have limited scope as they have focused on smaller datasets with large oil spills from exceptional disaster events [2]. This puts into question the suitability of such methods for building an automated system,” explains Yang.
“In contrast our method relies on large amounts of accessible data from Sentinel-1 and Artificial Intelligence to allow the detector to directly learn how each oil spill differs from their surroundings and possible look-alikes. This enables us to expand the scope to spills with different sizes and from a variety of sources. So far, we have reached an average precision for detecting all kinds of oil spills of 68 %.”
Study area with heat map of oil spills
The researchers at DLR are collaborating with the Research and Technology Centre Westcoast (FTZ) at Kiel University, Germany, and the Israeli Marine Data Centre (ISRAMAR) at the Israel Oceanographic & Limnological Research (IOLR).
The study is part of the binational DARTIS project, funded by the German Federal Ministry of Education and Research and the Israeli Ministry of Innovation, Science and Technology.
The teams are working on expanding the study area to both detect oil spills and estimate their trajectory, as well as including a simulation for an early warning system.
“Every time there is a Sentinel-1 acquisition, we go through the whole procedure to detect oil spills and then send this information to our partner team in Israel, who performs simulation of the oil slick trajectory. The whole procedure is automated, but before sending a warning alert to the authorities, we currently perform manual confirmation first,” concludes Yang.
The architecture is trained with information about look-alikes to further improve object detection and avoid false alarms in an alert detection system. The researchers are working on a detailed analysis to assess and improve the reliability of the system.
About the Copernicus Sentinels
The Copernicus Sentinels are a fleet of dedicated EU-owned satellites, designed to deliver the wealth of data and imagery that are central to the European Union's Copernicus environmental programme.
The European Commission leads and coordinates this programme, to improve the management of the environment, safeguarding lives every day. ESA is in charge of the space component, responsible for developing the family of Copernicus Sentinel satellites on behalf of the European Union and ensuring the flow of data for the Copernicus services, while the operations of the Copernicus Sentinels have been entrusted to ESA and EUMETSAT.
Did you know that?
Earth observation data from the Copernicus Sentinel satellites are fed into the Copernicus Services. First launched in 2012 with the Land Monitoring and Emergency Management services, these services provide free and open support, in six different thematic areas.
The Copernicus Marine Environment Monitoring Service (CMEMS) provides regular and systematic reference information on the physical and biogeochemical state, variability and dynamics of the ocean and marine ecosystems for the global ocean and the European regional seas.
References
[1] Y.-J. Yang et al. “A deep learning based oil spill detector using Sentinel-1 SAR imagery.” International Journal of Remote Sensing, 43:11, 4287-4314, (2022).
[2] S. Singha et al. “Satellite Oil Spill Detection Using Artificial Neural Networks.” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 6:6, (2013).