CNN Based Vehicle Track Detection in Coherent SAR Imagery: An Analysis of Data Augmentation

The coherence image as a product of a coherent SAR image pair can expose even subtle changes in the surface of a scene, such as vehicle tracks. For machine learning models, the large amount of required training data often is a crucial issue. A general solution for this is data augmentation. Standard techniques, however, were predominantly developed for optical imagery, thus do not account for SAR specific characteristics and thus are only partially applicable to SAR imagery. In this paper several data augmentation techniques are investigated for their performance impact regarding a CNN based vehicle track detection with the aim of generating an optimized data set. Quantitative results are shown on the performance comparison. Furthermore, the performance of the fully-augmented data set is put into relation to the training with a large non-augmented data set.