DUAL-POLARIMETRIC DECOMPOSITION OF SENTINEL-1 SAR IMAGE AND MACHINE LEARNING MODEL FOR OIL SPILL DETECTION: CASE OF MINDORO OIL SPILL

Abstract. Oil spills represent a significant environmental hazard necessitating timely detection to mitigate their detrimental effects. Synthetic Aperture Radar (SAR) technology serves as a remote sensing (RS)-based tool capable of detecting oil spills under varying weather conditions and at all times of day. SAR polarimetry, which assesses the polarization of the backscattered SAR signal, can effectively discriminate oil spills from other features that may manifest as dark regions in the SAR images. The integration of machine learning algorithms offers significant potential for enhancing the accuracy and efficiency of oil spill detection through SAR polarimetry. In recent years, several studies have introduced machine learning-based methodologies for this purpose, yet a comprehensive evaluation of their real-world performance remains essential. This study aimed to assess the efficacy of a machine learning (ML)-based approach for oil spill detection utilizing features derived from a dual-polarimetric decomposition method applied to Sentinel-1 SAR data. Results show that the machine learning-based approach achieved notable accuracy in oil spill detection reaching a score of 0.569 for intersection over union and 72.50 for f1-score of oil spill areas. Overall, this research underscores the potential of ML techniques as valuable tools for oil spill detection via SAR polarimetry.