A comparative study of fully automatic and semi-automatic methods for oil spill detection using Sentinel-1 data.

The oil spill detection and assessment study conducted in the Banten Province of Indonesia involves the application of Sentinel-1 satellite data and machine learning tools in the year 2024. Synthetic Aperture Radar (SAR) data were used with VV polarization to observe the surface characteristics, using an oil spill threshold of - 25 dB to differentiate clean water from the oil spill based on low backscatter intensity. After desiring image processing and binary masking applications on the data that improve visibility of the oil spill-affected zones, vectorization was conducted for integration into geographic information systems (GIS). A temporal analysis indicated high variability across the spill sizes with an extreme peak on May 16 (79.686 km²) and July 3 (41.593 km²), which are likely dictated by the weather and oceanographic conditions plus the ship traffic of that time. Wind pattern analysis via ERA5 reanalysis data presented more insight into spill dispersion dynamics. Three machine learning classifiers were applied toward oil spill detection, namely Artificial Neural Networks (ANN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). Performance metrics indicate the ANN outperformed in discriminative ability (AUC = 0.92), while RF was highly accurate (99.01%) and precise (99.02%). This clearly demonstrates the viability of using an integrated approach of remote sensing, advanced image processing, and supervised learning for environmental monitoring and provides important information for minimizing ecological impacts and optimizing disaster response plans for maritime areas. Such an integrated scheme calls for advanced technology to combat ecological threats in maritime areas and provides crucial evidence toward ongoing interventions to protect and manage marine ecosystems and the associated local communities.