Enhancing Cross Domain SAR Oil Spill Segmentation via Morphological Region Perturbation and Synthetic Label-to-SAR Generation

TL;DR

The paper tackles poor cross-domain generalization in SAR-based oil-spill segmentation by introducing MORP-Synth, a two-stage synthetic augmentation method that decouples geometry from texture. Stage A morphologically perturbs labeled slicks to create diverse shapes, while Stage B uses INADE-based synthesis to generate SAR-like textures conditioned on the edited masks, enabling effective mixed-domain training. Across seven segmentation architectures, MORP-Synth yields up to a 6 percentage-point gain in mean IoU on Peruvian data and substantial improvements for minority classes (oil and look-alike), particularly when combined with careful loss design and hard-negative mining. The work provides a practical, data-efficient pathway to improve coastal oil-spill monitoring in data-scarce regions, with implications for operational agencies and real-time surveillance.

Abstract

Deep learning models for SAR oil spill segmentation often fail to generalize across regions due to differences in sea-state, backscatter statistics, and slick morphology, a limitation that is particularly severe along the Peruvian coast where labeled Sentinel-1 data remain scarce. To address this problem, we propose \textbf{MORP--Synth}, a two-stage synthetic augmentation framework designed to improve transfer from Mediterranean to Peruvian conditions. Stage~A applies Morphological Region Perturbation, a curvature guided label space method that generates realistic geometric variations of oil and look-alike regions. Stage~B renders SAR-like textures from the edited masks using a conditional generative INADE model. We compile a Peruvian dataset of 2112 labeled 512$\times$512 patches from 40 Sentinel-1 scenes (2014--2024), harmonized with the Mediterranean CleanSeaNet benchmark, and evaluate seven segmentation architectures. Models pretrained on Mediterranean data degrade from 67.8\% to 51.8\% mIoU on the Peruvian domain; MORP--Synth improves performance up to +6 mIoU and boosts minority-class IoU (+10.8 oil, +14.6 look-alike).

Figures (11)

• Figure 1: Spatial distribution and volume classification of 66 oil spills recorded along the Peruvian coast from 2012 to 2022.
• Figure 2: Illustration of the MORP augmentation workflow for two representative spill regions. Each row shows the four-stage process: (a) original mask; (b) rotated/translated region with detected apices; (c) curvature-based edits (expansions in magenta; shrinkages in green); (d) final augmented mask. Additional examples in Appendix.
• Figure 3: Comparison of real and synthesized samples in the Mask-to-SAR generation stage. Columns: (a) real Sentinel-1 SAR patches; (b) ground-truth masks; (c) augmented masks produced by MORP; (d) synthetic SAR generated by INADE.
• Figure 4: Training curves on the Krestinin source domain (top row) and Peruvian target domain (bottom row). (a) Validation mIoU on Krestinin. (b) Validation loss on Krestinin. (c) Validation mIoU on Peru. (d) Validation loss on Peru.
• Figure 5: Comparison of oil spill segmentation predictions across five Peruvian Sentinel-1 patches. Columns show: (a) SAR input, (b) ground truth mask, (c–d) fine-tuned baselines (DeepLabV3+, SwinUnetTiny), and (e–f) fine-tuned models with synthetic data using $N{=}902$ and $N{=}1804$ real samples. Colors follow the palette defined in Sec. \ref{['sec:study-area-datasets']}).