Peruvian Coastal SAR Oil Spill Segmentation Dataset (Real, Synthetic, and Morphologically Perturbed Masks)

This dataset contains 2,112 labeled Sentinel-1 SAR image patches collected along the Peruvian coastline during documented oil spill events reported by OEFA and OSINERGMIN. Each patch includes pixel-level semantic annotations for five classes: sea, oil, look-alike phenomena, ships, and land. The dataset is designed to support research on oil spill detection, segmentation, and domain adaptation in maritime SAR imagery.To improve model generalization across geographic domains, the dataset also includes a set of morphologically perturbed masks generated using the MORP module (geometric transforms and curvature-based shape editing). These perturbations allow controlled modifications of oil spill boundaries, bulges, shrinks, and spatial configurations. In addition, a synthetic Label-to-SAR set is provided, where the perturbed masks are translated into realistic SAR textures using an INADE-based conditional generation model. This enables researchers to train segmentation architectures using mixed real + synthetic samples without requiring additional SAR acquisitions.The dataset supports benchmarking of deep learning methods for oil spill segmentation, domain adaptation (Mediterranean \u2192 Peru), and robustness to spill morphology variations. Initial experiments using a ResNet-DeepLabv3+ architecture demonstrate performance improvements of +6.00 pp mIoU, +10.81 pp oil IoU, and +14.6 pp look-alike IoU when including MORP-Synth data during training.This dataset aims to facilitate reproducibility and standardized evaluation for SAR-based maritime environmental monitoring, oil spill response, and remote sensing research focused on cross-domain generalization.