Realistic augmentations for SAR images of ships for machine learning

Abstract:
The majority of work on machine learning for classification is done using optical images, but work on SAR images is also appearing. One challenge with SAR as opposed to optical images is the lack of real data for training. Several solutions have been suggested to overcome this problem, like using simulations and augmentations. Many augmentation techniques used for optical imagery are not directly applicable to SAR imagery. Some of them may still be useful technically for machine learning, but may not make much sense physically. A collection of augmentations for SAR images of ships in particular is proposed here.