SGDiff-OS: Synthetic SAR Oil Spill Backscatter and Mask Pairs

SGDiff-OS is a synthetic dataset for marine oil spill remote sensing, containing generated pairs of SAR backscatter coefficient (σ⁰) fields and corresponding binary oil spill mask labels. The dataset was created by author Rui Zhang and published on figshare in May 2026. It was constructed to address the limited availability of well-annotated SAR oil spill samples with structural consistency and scattering representation.

Use Cases

• Augmenting training data for oil spill segmentation models based on generated mask and backscatter pairs.
• Training or evaluating semantic-guided diffusion models for SAR data generation based on spotted, blocky, and slender oil morphologies.
• Analyzing statistical properties of synthetic SAR backscatter coefficients based on the reported mean difference within 2 dB between oil and sea regions.
• Benchmarking downstream segmentation task performance based on the reported mIoU of 0.7617 and F1-score of 0.8606.
• Studying the contrast relationship between oil and water in SAR imagery based on the generated scattering characteristics.

Strengths

• Generated samples achieve high visual and statistical consistency with real data, with a mean σ⁰ difference between oil and clean sea regions controlled within 2 dB.
• Incorporating generated samples led to a reported maximum improvement of 26.0% in mIoU for segmentation under limited training data.
• The model incorporates semantic modulation for prior knowledge of oil spill morphologies (spotted, blocky, slender), improving structural representation.

Limitations

• Dataset size is only 155.5 KB, indicating a very limited scope and number of generated samples.
• Row count and specific column definitions are unknown, requiring manual inspection after download.
• The data is synthetic, generated by the SGDiff-OS model, and may not capture the full complexity of real-world SAR oil spill scenarios.

Provenance

Source

Rui Zhang

Collection Method

Generated by the Semantic-Guided Diffusion for SAR Oil Spill Generation (SGDiff-OS) model trained on a self-constructed SAR oil spill dataset.

Freshness

Last updated 2026-05-13 08:57:02.