Domain fusion based feature extraction for SAR ATR

The all-weather and light condition operability of synthetic aperture radar (SAR) imaging systems makes them the optimal choice for several civilian and military remote sensing applications. Deep learning methods have demonstrated state-ofthe-art classification performance on standard SAR datasets such as the Moving and Stationary Target Acquisition and Recognition (MSTAR) Standard Operating Conditions (SOC) 10-target dataset. However, high acquisition costs limit the availability SAR domain data, both in number and diversity for use in training neural networks. This in turn limits the performance of these networks when used to classify SAR images acquired using radar system specifications and imaging environments that differ from the specifications used to create the images used for training. In this work, Siamese Networks, made up of twin AlexNet-based CNNs, were trained using subsets of the Military Ground Target Dataset (MGTD) and MSTAR datasets to learn radar specification and imaging environment invariant features thereby increasing the classification performance on the MGTD test set by 4.17%.