A Two-Stage Deep Learning Method for Marine Oil Spill Localization and Segmentation From Synthetic Aperture Radar Images

Oil spills have devastating impacts on ocean ecosystems, leading to significant economic losses. Consequently, it is crucial to develop effective techniques for the rapid detection of marine oil spills using satellite observations. Existing methods for detecting oil spills using synthetic aperture radar (SAR) often face challenges related to algorithm complexity, limited data availability, and practical implementation. To address these issues, we propose a two-stage deep learning (DL) method to automatically localize and segment oil spills in Sentinel-1 SAR images. First, we utilize the Faster R-CNN model to identify oil spill locations. Subsequently, sub-images containing the oil spills are input into the UNet++ model to achieve pixel-wise segmentation. We have created a comprehensive dataset consisting of 3,166 well-annotated oil spill samples derived from 1,000 Sentinel-1 VV-polarized SAR images. This unique dataset will be used for training, validation, and testing the two DL models. Experimental results indicate that the precision and recall for oil spill localization are 87.46% and 87.59%, respectively, while for oil spill segmentation, they are 89.55% and 89.37%. The proposed method demonstrates optimal oil detection performance at wind speeds between 2 m/s and 10 m/s. Furthermore, the total runtime for oil spill localization and segmentation is less than 1 minute for each full-resolution SAR image. These results suggest that the proposed approach has strong potential for operational oil spill monitoring using spaceborne SAR images.