Quad‑polarization synthetic aperture radar dataset for marine oil spill and look‑alike segmentation

The quad‑polarization oil spill detection dataset is a collection of fully polarimetric L‑band synthetic aperture radar scenes acquired by the NASA/JPL uninhabited aerial vehicle synthetic aperture radar (UAVSAR) over the U.S. Gulf of Mexico from 2010 to 2022. The dataset comprises 41 multi‑look complex SAR scenes with a spatial resolution of roughly 6m×4m and incidence angles between 27° and 67°. For each scene, a 3 × 3 polarimetric coherence matrix is computed using the quad polarization channels, (HH, HV, VH, VV) namely horizontal transmit horizontal receive, horizontal transmit vertical receive, vertical transmit horizontal receive, and vertical transmit vertical receive. This matrix is then represented as a nine channel real valued feature set. Pixel‑level segmentation masks annotate four classes: sea surface, oil spill, look‑alike phenomena and land with more than 80 million labelled pixels. Ground‑truth labels were generated with a semi‑automated workflow combining manual labelling, Random Forest‑based label propagation and manual refinement. QPOSD is intended to enable reproducible research, benchmarking and polarimetric feature analysis for SAR‑based oil spill and look‑alike segmentation. The dataset is accessible at DOI: https://doi.org/10.5281/zenodo.19258036 .