Satellite-based oil spill detection using Synthetic Aperture Radar (SAR) imagery is a fundamental tool for emergency response and the mitigation of marine environmental disasters. However, traditional operational methods based on backscatter intensity thresholds (using GRD products) exhibit a high false alarm rate due to lookalike phenomena, such as biomass accumulations or low-wind roughness areas. Although complex polarimetry (SLC products) provides an accurate physical characterization of scattering mechanisms that resolves this ambiguity, its high computational requirements and limited availability restrict its application in continuous monitoring.
This research evaluates the feasibility and effectiveness of integrating pseudo-polarimetric parameters (pseudo-entropy Hᶜ, pseudo-angle θᶜ, and pseudo-purity mᶜ), mathematically derived from Sentinel-1 dual-polarization GRD products, into Machine Learning frameworks to optimize oil spill detection. The methodological workflow encompasses the validation of the physical separability of the target classes (Oil Spill, Water, and Lookalike), followed by a comparative performance assessment of different feature sets (standard intensity versus pseudo-polarimetry) using both unsupervised clustering techniques (K-Means) and supervised multiclass classification.
Preliminary results suggest a dual behavior. While pure intensity (VV band) appears to dominate the accuracy of local geometric segmentation due to its high radiometric contrast, the integration of pseudo-polarimetric parameters is expected to play a decisive role in the global supervised domain by reducing the confusion between oil spills and heterogeneous false alarms. Based on this premise, the study proposes evaluating a two-stage hybrid processing pipeline in which pseudo-polarimetry drives the initial discrimination of anomalous regions, while intensity data refine their spatial delineation. This approach is expected to demonstrate that it is possible to achieve a level of physical robustness comparable to that of complex polarimetric data without sacrificing the operational efficiency of amplitude-based GRD products.
Co-author: Alejandro Tellez Quiñones (atellez@centrogeo.edu.mx)
