Periodic Reporting for period 1 - SOSeas (Searching for Oil Spills on Sea Surfaces)
Okres sprawozdawczy: 2023-09-01 do 2026-06-30
Podsumowanie kontekstu i ogólnych celów projektu
Marine oil spills can cause severe environmental and socioeconomic impacts, requiring rapid access to reliable information to support monitoring and emergency response. Because oil spills can spread rapidly over large areas and occur under a wide range of weather and sea-state conditions, satellite Earth Observation (EO), particularly Synthetic Aperture Radar (SAR), provides a strategic capability for systematic marine pollution monitoring. SAR can observe large ocean areas independently of daylight and under most weather conditions and is therefore widely used by operational services for oil spill detection. However, current operational SAR-based monitoring primarily provides information on the location, shape and extent of oil slicks, while information on their internal characteristics and relative oil thickness is not yet routinely available. This represents an important operational gap because clean-up actions are more effective when directed towards relatively thicker oil accumulations, commonly referred to as “actionable oil”.
The Searching for Oil Spills on Sea Surfaces (SOSeas) project was designed to address this gap by integrating Sentinel-1 SAR observations, artificial intelligence (AI), data science and operational oil spill monitoring and response knowledge. Its overall objective is to co-develop an AI-based framework for oil spill detection and characterization, advancing from conventional slick detection and delineation towards the extraction of value-added information on relative oil thickness. The long-term objective is to support the identification of potentially actionable oil and thereby contribute to more informed decision-making during marine pollution emergencies.
The scientific pathway was structured through four interconnected work packages. WP1 focused on building the SAR oil spill dataset and the processing framework required to generate consistent SAR-derived products. WP2 incorporated controlled oil spill samples to support comparative analyses under different oil and sea state conditions. WP3 focused on developing and validating approaches for oil spill detection and characterization using statistical analysis, machine learning and deep learning. WP4 was conceived to translate these scientific developments into a user-oriented prototype interface and explore their potential operational application.
A fundamental challenge encountered during the project was the scarcity of sufficiently large and validated oil spill datasets suitable for AI development and, particularly, relative oil-thickness research. Existing open oil spill datasets mainly provide processed NRCS imagery and were not specifically designed to preserve the SAR information required to investigate subtle intra-slick variations. SOSeas therefore substantially expanded its original data strategy and developed a large-scale transoceanic benchmark comprising approximately 1,700 field- and expert-validated oil spill samples from multiple oceanic regions and operational sources, substantially increasing the scale and diversity of the empirical basis available for SAR-based oil spill research. Rather than providing only conventional NRCS imagery, the SOSeas framework preserves calibrated SAR measurements and integrates multiple physically meaningful SAR-derived products, signal-quality indicators and ancillary environmental information.
This expanded approach transformed an initial data limitation into one of the project's main scientific outcomes. The resulting SOSeas.Dataset provides the foundation for investigating oil spill characteristics across different geographic regions, sea states, acquisition geometries and spill conditions, while supporting the development and validation of scalable AI approaches. Particular emphasis is placed on the Damping Ratio (DR), a SAR-derived contrast metric investigated within SOSeas for both oil detection and relative oil-thickness characterization.The SOSeas-Bonn proof-of-concept validated this methodological approach and established the basis for its extension to the complete transoceanic benchmark .
The project pathway to impact therefore extends from data and methodological development to scientific knowledge and, ultimately, operational application. In the short term, SOSeas provides new benchmark data, validated processing methodologies and scientific results that can support further research in SAR, AI and marine pollution monitoring. In the medium and longer term, these developments provide the scientific basis for thickness-aware AI models and thematic products capable of identifying areas with a greater likelihood of relatively thicker oil accumulation. Such information could complement existing operational oil spill detection products and support responders in prioritizing surveillance and response actions, potentially improving oil recovery efficiency while reducing environmental and socioeconomic impacts.
Through this pathway, SOSeas contributes to bridging the gap between Earth Observation research and operational marine pollution monitoring and response, while supporting broader objectives related to ocean protection, disaster response and sustainable ocean governance. Its scientific outputs and continued engagement with research institutions, governmental organizations and operational marine pollution monitoring stakeholders provide a pathway for the results to be further validated, exploited and translated into future operational applications.
The Searching for Oil Spills on Sea Surfaces (SOSeas) project was designed to address this gap by integrating Sentinel-1 SAR observations, artificial intelligence (AI), data science and operational oil spill monitoring and response knowledge. Its overall objective is to co-develop an AI-based framework for oil spill detection and characterization, advancing from conventional slick detection and delineation towards the extraction of value-added information on relative oil thickness. The long-term objective is to support the identification of potentially actionable oil and thereby contribute to more informed decision-making during marine pollution emergencies.
The scientific pathway was structured through four interconnected work packages. WP1 focused on building the SAR oil spill dataset and the processing framework required to generate consistent SAR-derived products. WP2 incorporated controlled oil spill samples to support comparative analyses under different oil and sea state conditions. WP3 focused on developing and validating approaches for oil spill detection and characterization using statistical analysis, machine learning and deep learning. WP4 was conceived to translate these scientific developments into a user-oriented prototype interface and explore their potential operational application.
A fundamental challenge encountered during the project was the scarcity of sufficiently large and validated oil spill datasets suitable for AI development and, particularly, relative oil-thickness research. Existing open oil spill datasets mainly provide processed NRCS imagery and were not specifically designed to preserve the SAR information required to investigate subtle intra-slick variations. SOSeas therefore substantially expanded its original data strategy and developed a large-scale transoceanic benchmark comprising approximately 1,700 field- and expert-validated oil spill samples from multiple oceanic regions and operational sources, substantially increasing the scale and diversity of the empirical basis available for SAR-based oil spill research. Rather than providing only conventional NRCS imagery, the SOSeas framework preserves calibrated SAR measurements and integrates multiple physically meaningful SAR-derived products, signal-quality indicators and ancillary environmental information.
This expanded approach transformed an initial data limitation into one of the project's main scientific outcomes. The resulting SOSeas.Dataset provides the foundation for investigating oil spill characteristics across different geographic regions, sea states, acquisition geometries and spill conditions, while supporting the development and validation of scalable AI approaches. Particular emphasis is placed on the Damping Ratio (DR), a SAR-derived contrast metric investigated within SOSeas for both oil detection and relative oil-thickness characterization.The SOSeas-Bonn proof-of-concept validated this methodological approach and established the basis for its extension to the complete transoceanic benchmark .
The project pathway to impact therefore extends from data and methodological development to scientific knowledge and, ultimately, operational application. In the short term, SOSeas provides new benchmark data, validated processing methodologies and scientific results that can support further research in SAR, AI and marine pollution monitoring. In the medium and longer term, these developments provide the scientific basis for thickness-aware AI models and thematic products capable of identifying areas with a greater likelihood of relatively thicker oil accumulation. Such information could complement existing operational oil spill detection products and support responders in prioritizing surveillance and response actions, potentially improving oil recovery efficiency while reducing environmental and socioeconomic impacts.
Through this pathway, SOSeas contributes to bridging the gap between Earth Observation research and operational marine pollution monitoring and response, while supporting broader objectives related to ocean protection, disaster response and sustainable ocean governance. Its scientific outputs and continued engagement with research institutions, governmental organizations and operational marine pollution monitoring stakeholders provide a pathway for the results to be further validated, exploited and translated into future operational applications.
Prace wykonane od początku projektu do końca okresu sprawozdawczego oraz najważniejsze dotychczasowe rezultaty
SOSeas established a complete methodological framework for large-scale SAR-based oil spill detection and relative oil-thickness characterization. The work included the prospection and validation of real oil spill events, Sentinel-1 data retrieval and processing, generation of physically meaningful SAR-derived products, design of binary oil-spill masks, statistical analyses, and AI-based experiments.
A proof-of-concept benchmark, SOSeas-Bonn, was first developed using field-validated oil spills from the Bonn Agreement. It was used to establish and validate the complete processing pipeline before scaling the methodology to the transoceanic SOSeas.Dataset. The framework preserves calibrated Sentinel-1 measurements and integrates Damping Ratio (DR), incidence angle, sensor-noise and signal-quality information, together with ancillary wind information when available. The validated approach was subsequently scaled to approximately 1,700 field- and expert-validated oil spill samples, providing a substantially expanded empirical basis for large-scale SAR analysis.
Large-scale statistical analyses were conducted to investigate the physical behaviour of SAR measurements over polluted and non-oiled waters under varying acquisition geometry, wind and sensor-noise conditions. The results demonstrated that DR reduces the influence of these factors over non-oiled surfaces while improving oil-to-background separability and preserving intra-slick variability compared with conventional NRCS.
Controlled deep-learning experiments were subsequently performed using NRCS and DR as inputs for oil spill segmentation. DR achieved improved oil-spill detection performance, demonstrating that its physical and statistical advantages can also translate into improved AI-based segmentation. These experiments established and validated the methodological basis for scaling the detection framework to the complete SOSeas.Dataset and for testing additional deep-learning architectures.
The methodology was further extended towards relative oil-thickness characterization. DR-derived categories were used to investigate intra-slick variability and the spatial distribution of relatively thicker oil accumulations across independent datasets and different oceanic regions. The analyses revealed recurrent relative-thickness distribution patterns across geographically and operationally distinct datasets, providing evidence that DR captures meaningful intra-slick structures and supporting its potential use for thematic oil slick characterization.
A proof-of-concept benchmark, SOSeas-Bonn, was first developed using field-validated oil spills from the Bonn Agreement. It was used to establish and validate the complete processing pipeline before scaling the methodology to the transoceanic SOSeas.Dataset. The framework preserves calibrated Sentinel-1 measurements and integrates Damping Ratio (DR), incidence angle, sensor-noise and signal-quality information, together with ancillary wind information when available. The validated approach was subsequently scaled to approximately 1,700 field- and expert-validated oil spill samples, providing a substantially expanded empirical basis for large-scale SAR analysis.
Large-scale statistical analyses were conducted to investigate the physical behaviour of SAR measurements over polluted and non-oiled waters under varying acquisition geometry, wind and sensor-noise conditions. The results demonstrated that DR reduces the influence of these factors over non-oiled surfaces while improving oil-to-background separability and preserving intra-slick variability compared with conventional NRCS.
Controlled deep-learning experiments were subsequently performed using NRCS and DR as inputs for oil spill segmentation. DR achieved improved oil-spill detection performance, demonstrating that its physical and statistical advantages can also translate into improved AI-based segmentation. These experiments established and validated the methodological basis for scaling the detection framework to the complete SOSeas.Dataset and for testing additional deep-learning architectures.
The methodology was further extended towards relative oil-thickness characterization. DR-derived categories were used to investigate intra-slick variability and the spatial distribution of relatively thicker oil accumulations across independent datasets and different oceanic regions. The analyses revealed recurrent relative-thickness distribution patterns across geographically and operationally distinct datasets, providing evidence that DR captures meaningful intra-slick structures and supporting its potential use for thematic oil slick characterization.
Innowacyjność oraz oczekiwany potencjalny wpływ (w tym dotychczasowe znaczenie społeczno-gospodarcze i szersze implikacje społeczne projektu)
SOSeas advances beyond the state of the art by establishing a new generation of SAR benchmarks designed not only for oil spill detection, but also for relative oil-thickness characterization and different maritime applications. Existing oil spill benchmarks provide primarily NRCS imagery and are generally distributed after extensive preprocessing, such as orthorectification, filtering, resampling, normalization and, in some cases, data compression, which can modify the radiometric and spatial characteristics of the original SAR measurements and limit their suitability for investigating subtle intra-slick backscatter variability. To overcome this limitation, SOSeas.Dataset preserves calibrated Sentinel-1 measurements and integrates them within a multi-layer benchmark containing physically meaningful SAR-derived products, signal-quality indicators and ancillary environmental information, providing a substantially richer basis for both physical analyses and AI development.
Using the SOSeas-Bonn proof-of-concept benchmark, the project demonstrated at large scale that Damping Ratio (DR) provides a more robust representation of oil-covered surfaces than conventional NRCS. DR reduces the influence of incidence angle, wind variability and sensor noise over non-oiled surfaces while enhancing oil-to-background contrast and preserving intra-slick variability. Global overlap between polluted and non-oiled pixels decreased from 35.5% with NRCS to 14.0% with DR, demonstrating substantially improved class separability.
SOSeas also introduced and systematically evaluated DR as an input feature for large-scale deep-learning-based oil spill detection using Sentinel-1 data. Controlled semantic-segmentation experiments showed higher Recall, F1-score and oil-spill Intersection-over-Union with DR than with conventional NRCS, demonstrating that its physical and statistical advantages translate into improved AI-based detection performance.
Building on these findings, SOSeas extended the framework towards thematic characterization of relative oil thickness. The transoceanic scale of SOSeas.Dataset bringing together validated oil spills from different oceanic regions and environmental conditions, enables relative oil-thickness patterns to be systematically investigated across geographically and operationally diverse Sentinel-1 observations. Preliminary analyses across independent datasets reveal recurrent spatial patterns, with higher DR categories, associated with a greater likelihood of relatively thicker oil accumulation, consistently occupying smaller portions of the slick.
Together, these advances establish SOSeas.Dataset as a unique scientific resource extending beyond oil spill applications. By preserving calibrated SAR measurements alongside multiple derived products, the benchmark can support the development of next-generation AI-based solutions for marine applications, as well as research on SAR processing, feature extraction and computer-vision methodologies. Most importantly, the large-scale validation of DR and the emerging transoceanic relative-thickness patterns provide the scientific knowledge required to progress from conventional oil spill detection towards the future operational integration of relative oil-thickness characterization into marine pollution monitoring and emergency-response services.
Further uptake will require completing the large-scale validation of the methodology across the full SOSeas.Dataset extending and comparing AI architectures, and further validating the relative-thickness characterization under diverse environmental and operational conditions. Operational demonstration and continued engagement with marine pollution monitoring and response organizations will be particularly important to assess how thickness-aware products can be integrated into existing surveillance and emergency-response workflows.
Using the SOSeas-Bonn proof-of-concept benchmark, the project demonstrated at large scale that Damping Ratio (DR) provides a more robust representation of oil-covered surfaces than conventional NRCS. DR reduces the influence of incidence angle, wind variability and sensor noise over non-oiled surfaces while enhancing oil-to-background contrast and preserving intra-slick variability. Global overlap between polluted and non-oiled pixels decreased from 35.5% with NRCS to 14.0% with DR, demonstrating substantially improved class separability.
SOSeas also introduced and systematically evaluated DR as an input feature for large-scale deep-learning-based oil spill detection using Sentinel-1 data. Controlled semantic-segmentation experiments showed higher Recall, F1-score and oil-spill Intersection-over-Union with DR than with conventional NRCS, demonstrating that its physical and statistical advantages translate into improved AI-based detection performance.
Building on these findings, SOSeas extended the framework towards thematic characterization of relative oil thickness. The transoceanic scale of SOSeas.Dataset bringing together validated oil spills from different oceanic regions and environmental conditions, enables relative oil-thickness patterns to be systematically investigated across geographically and operationally diverse Sentinel-1 observations. Preliminary analyses across independent datasets reveal recurrent spatial patterns, with higher DR categories, associated with a greater likelihood of relatively thicker oil accumulation, consistently occupying smaller portions of the slick.
Together, these advances establish SOSeas.Dataset as a unique scientific resource extending beyond oil spill applications. By preserving calibrated SAR measurements alongside multiple derived products, the benchmark can support the development of next-generation AI-based solutions for marine applications, as well as research on SAR processing, feature extraction and computer-vision methodologies. Most importantly, the large-scale validation of DR and the emerging transoceanic relative-thickness patterns provide the scientific knowledge required to progress from conventional oil spill detection towards the future operational integration of relative oil-thickness characterization into marine pollution monitoring and emergency-response services.
Further uptake will require completing the large-scale validation of the methodology across the full SOSeas.Dataset extending and comparing AI architectures, and further validating the relative-thickness characterization under diverse environmental and operational conditions. Operational demonstration and continued engagement with marine pollution monitoring and response organizations will be particularly important to assess how thickness-aware products can be integrated into existing surveillance and emergency-response workflows.