SAR-based Oil Spill Detection and Characterization using Damping Ratio and Semantic Segmentation to Advance Operational AI-based solutions for Oceanic Monitoring and Protection

The development of scalable AI-based solutions for oil spill detection, discrimination, and characterization is crucial for advancing marine protection, disaster response, and sustainable ocean governance. Due to the attenuation of sea surface roughness, oil spills are detected as low backscattering regions in Synthetic Aperture Radar (SAR) imagery, being a strategic data source for operational services dedicated to marine pollution monitoring. Near-real-time information on the location, extent, and shape of oil-covered areas represents the primary SAR-derived input for oil spill response (OSR). In a subsequent stage, characterizing relative oil thickness variations within slicks becomes critical for improving cleanup effectiveness, which is higher over thicker oil layers commonly referred to as “actionable oil”.

A strong contrast between dark features and the surrounding sea-surface is required for reliable detection and represents a key property for improving discrimination and characterization using data-driven approaches. In this context, the Damping Ratio (DR) has been demonstrated as a meaningful SAR-based feature for enhancing sea surface contrast. Compared to the Normalized Radar Cross Section (NRCS), DR is less affected by incidence angle and wind intensity, offering strong potential for the development of robust, operational AI-based systems for oil slick detection and characterization.

Existing deep learning approaches for oil spill detection rely exclusively on NRCS, which has shaped available SAR datasets toward pre-processed products unsuitable for oil slick characterization. The project “Searching for Oil Spills on Sea Surfaces” (SOSeas) proposes a two-stage AI-based framework, in which deep learning is used for oil spill detection and delineation, followed by the thematic characterization of relative oil thickness within intra-slicks, both utilizing DR as the primary feature. Achieving this objective required the construction of the SOSeas.Dataset, a new, large-scale, field-validated oil spill benchmark that goes beyond NRCS, providing SAR-derived products from Sentinel-1, especially DR, in raw format to preserve the sea-surface backscattering properties important for characterization.

A proof-of-concept dataset comprising 143 oil spills, field-validated by the Bonn Agreement and primarily located in the North Sea, was used to train, test, and validate a UNet–based semantic segmentation model for oil spill detection. Two identically configured models were trained using either DR or NRCS as input to directly compare their detection performance. To evaluate the discriminative power of DR versus NRCS, binary oil spill masks were used as labels to distinguish polluted water (PW) from non-oiled (NO) areas, which include clean ocean and lookalikes. Models trained with DR consistently outperformed NRCS-based models, achieving higher Intersection over Union (NRCS: 0.4058; DR: 0.5491) and F1-scores (NRCS: 0.58; DR: 0.71) for the polluted water class.

The validation of the DR as a better feature for oil detection lays the foundation for developing the second stage as a future perspective, integrating DR and deep learning for oil slick characterization. Finally, the SOSeas.Dataset lays the groundwork for developing new AI-driven solutions capable of processing large volumes of SAR data, identifying patterns, and extracting useful information in near-real time, supporting operational agencies while enhancing monitoring and OSR actions for ocean protection.