An improved Pix2PixGAN sample generation method for SAR oil spill detection

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Marine Science

Background:

• Marine oil spills present significant ecological and socio-economic risks.
• Limited annotated Synthetic Aperture Radar (SAR) oil-spill data hinders deep learning model development.

Purpose of the Study:

• To develop an improved Pix2PixGAN for generating realistic SAR oil-spill samples.
• To enhance the training and generalization capabilities of deep learning-based oil-spill detection models.

Main Methods:

• Proposed an improved Pix2PixGAN using real SAR images as background priors.
• Integrated oil-spill masks with multi-channel random noise for constrained generation.
• Incorporated spatially adaptive normalization, PatchGAN, least-squares adversarial loss, Dropout, and segmentation-consistency constraints.

Main Results:

• The proposed method outperformed existing GANs (Pix2Pix, PGGAN, BEGAN) in key metrics (FID, KID, Global BIHD, ENL, GMHD).
• Generated samples demonstrated high consistency with real SAR data in feature distribution and statistical characteristics.
• Models trained with generated samples achieved performance comparable to those trained with real samples, proving sample usability.

Conclusions:

• The generated SAR oil-spill samples contain valuable structural and textural information.
• The synthetic data can effectively supplement limited real annotated samples for training deep learning models.
• The approach shows promise for improving SAR oil-spill detection in data-scarce scenarios.