Automating Oil Spill Segmentation in Synthetic Aperture Radar Images

Marine oil spills at sea may cause disastrous consequences for the marine environment, especially the marine ecosystem. Satellite-based synthetic aperture radar (SAR) provides an important means for monitoring marine oil spills due to its advantages of all-weather and all-time observation ability. Thus, to make a timely damage assessment and spread control of oil spills, it is vital to accurately segment oil spills through SAR images. In oil spill SAR image segmentation, several inevitable challenges exist, and these challenges mainly include: 1) oil spills in SAR images normally exhibit various areas with irregular shapes, and this makes difficulty for accurately segmenting oil spill areas in SAR images. 2) look-alikes can be confused with oil spills and thus being a significant obstacle for segmenting real oil spill areas in SAR images. 3) the availability of oil spill SAR image data is limited in practice, and this makes difficulty in training a qualified segmentation model. To address these challenges, in this thesis, I firstly propose a novel marine oil spill SAR image segmentation method by considering the physics of SAR imagery and oil spill segmentation simultaneously to perform accurate oil spill segmentation. Secondly, I incorporate the intrinsic distribution of backscatter values in SAR images to construct a deep neural segmentation network, which is acpable of segmentation oil spill SAR images include look-alikes.

Finally, I construct a novel oil spill SAR image segmentation network by leveraging the seminal representation of SAR images and the training for oil spill segmentation comprehensively, which operates oil spill segmentation with a small amount of training data. Experimental evaluations over different types of oil spill SAR images demonstrate the proposed methods achieve an average of around 10% improvement for oil spill SAR image segmentation.