Research on high-spectral polarization detection and classification of submerged oil based on multi-dimensional information

This study tackles the problem of detecting submerged oil in underwater environments. We developed a hyperspectral–polarimetric imaging system that spans 500–900 nm with 201 spectral bands and an average spectral resolution better than 3 nm, and we established a complete denoising–fusion–classification workflow to improve the separability and recognition reliability of submerged oil in complex waters. Samples were collected in real seawater, and a dataset was built that includes multiple oil types and look‑alike materials (e.g., plastics, driftwood) to approximate actual underwater spill conditions. Four linear‑polarization orientations were acquired and used to compute the Degree of Linear Polarization and the Stokes parameters. In preprocessing, we compared several filtering methods—Non‑Local Means filtering, Bilateral Filter, Guided Filter, Wiener filter, and Adaptive Median Filter—to suppress noise. In the fusion stage, we evaluated Principal Component Analysis, wavelet transform, Laplacian pyramid, and two rules of the Non‑Subsampled Contourlet Transform to achieve structural preservation and detail enhancement across modalities; for classification, an artificial neural network was used for supervised learning and recognition. The results show that the combination of Guided Filter and NSCT Rule 2 delivers the best performance (overall accuracy 98.09 %, Kappa 0.9767). The linear correlation co- efficient in the repeated experiment reached 0.971, demonstrating the method’s effectiveness and robustness in real underwater scenarios. These findings indicate that the proposed hyperspectral–polarization fusion imaging scheme not only significantly improves the identification accuracy of submerged oil, but also offers solid engineering potential, providing a new technical approach and effective tool for marine oil‑pollution monitoring.