SAR Specific Noise Based Data Augmentation for Deep Learning

Deep learning techniques provide a significant performance increase in automatic target recognition (ATR) of synthetic aperture radar (SAR) images. Due to acquisition complexity, SAR images are often scarce and the performance of deep learning methods is strongly affected by the lack of diversity and low number of images available for training. Data augmentation is a solution that tackles the problem of reduced data by artificially expanding a training set. In this paper, we propose a data augmentation solution that adds Weibull noise to the High Range Resolution Profiles before SAR processing. The resulting noisy images are added to the original training set. A standard CNN is used to evaluate the impact of the proposed data augmentation on the Cranfield University Military Ground Target Dataset (MGTD). The analysis of performance shows are compared with those obtained with the classic translation data augmentation. Results show a 91% correct classification rate is achieved when a combination of the translation and the Weibull noise data augmentation is employed, compared to 86% with a classic translation data augmentation alone and 77% on standard images without data augmentation.