Augmentation of Maritime SAR Imagery using Realistic Motion Models

The use of convolutional neural networks (CNNs) for automatic detection and recognition of ships using synthetic aperture radar (SAR) imagery can be challenging as the amount of labelled training data available is often insufficient. Techniques such as transfer learning and data augmentation are therefore required to achieve sufficient classification performance. This paper addresses the latter technique by transforming images of focussed ships into rotated and blurred versions based on realistic motion. Two approaches are demonstrated with motion added to the entire SAR image and then to the dominant scatterers extracted using the CLEAN technique. The technique works on both real and complex data and is demonstrated using data from ICEYE's extended spotlight SAR mode.