SGDiff-OS: a semantic-guided diffusion model for SAR oil spill scattering and mask field generation

Marine oil spill remote sensing relies heavily on high-quality annotated data, yet the limited availability of well-annotated SAR oil spill samples with backscatter information and structural labels remains a challenge. To address the lack of structural consistency and scattering representation of existing generation methods, this study proposes a Semantic-Guided Diffusion for SAR Oil Spill Generation model (SGDiff-OS) for the joint generation of SAR oil spill backscatter coefficient σ° and corresponding binary oil spill labels, trained on a self-constructed SAR oil spill dataset. The model adopts a dual-channel conditional diffusion framework and incorporates semantic modulation to encode prior knowledge of spotted, blocky and slender morphologies, thereby improving structural representation and category controllability. Experimental results demonstrate that the generated samples show general agreement with real data in visual quality and overall statistical trends, although measurable distributional differences remain. The global mean σ° differs by approximately 1.96 dB between the real and generated samples. The generated samples retain the lower-backscatter relationship of oil-covered regions relative to seawater, although the oil-water contrast decreases from 6.59 dB to 4.65 dB. Furthermore, in an expert annotation evaluation of generated image mask pairs, the generated masks achieve an mIoU of 0.7658 and an F1-score of 0.8637 when evaluated against reference annotations. Under limited training data conditions, incorporating generated samples leads to a maximum improvement of 24.2% in mIoU, suggesting its potential value for sample augmentation under limited training conditions. These results demonstrate that the proposed model can generate SAR oil spill samples with the dominant scattering trends of real SAR data, providing supplementary data for oil spill detection and segmentation tasks.

Authors

Publication Details

Journal
Figshare
Published
2026-11-18
DOI
https://doi.org/10.6084/m9.figshare.35046619
Primary Topic
Oil Spill Detection and Mitigation
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

SGDiff-OS: a semantic-guided diffusion model for SAR oil spill scattering and mask field generation

Dongmei Song, Weizeng Shao, Huiqiao Wang, Rui Zhang et al.
4 links
Figshare
Oil Spill Detection and Mitigation
article

SGDiff-OS: a semantic-guided diffusion model for SAR oil spill scattering and mask field generation

Dongmei Song, Weizeng Shao, Huiqiao Wang, Rui Zhang, Yang Song, Yong Wan, Weibo Zhang
article en

Abstract

Marine oil spill remote sensing relies heavily on high-quality annotated data, yet the limited availability of well-annotated SAR oil spill samples with backscatter information and structural labels remains a challenge. To address the lack of structural consistency and scattering representation of existing generation methods, this study proposes a Semantic-Guided Diffusion for SAR Oil Spill Generation model (SGDiff-OS) for the joint generation of SAR oil spill backscatter coefficient σ° and corresponding binary oil spill labels, trained on a self-constructed SAR oil spill dataset. The model adopts a dual-channel conditional diffusion framework and incorporates semantic modulation to encode prior knowledge of spotted, blocky and slender morphologies, thereby improving structural representation and category controllability. Experimental results demonstrate that the generated samples show general agreement with real data in visual quality and overall statistical trends, although measurable distributional differences remain. The global mean σ° differs by approximately 1.96 dB between the real and generated samples. The generated samples retain the lower-backscatter relationship of oil-covered regions relative to seawater, although the oil-water contrast decreases from 6.59 dB to 4.65 dB. Furthermore, in an expert annotation evaluation of generated image mask pairs, the generated masks achieve an mIoU of 0.7658 and an F1-score of 0.8637 when evaluated against reference annotations. Under limited training data conditions, incorporating generated samples leads to a maximum improvement of 24.2% in mIoU, suggesting its potential value for sample augmentation under limited training conditions. These results demonstrate that the proposed model can generate SAR oil spill samples with the dominant scattering trends of real SAR data, providing supplementary data for oil spill detection and segmentation tasks.

Figshare
Oil Spill Detection and Mitigation
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon