Advances in the development of deep neural networks and other machine learning (ML) algorithms, combined with ever more powerful hardware and the huge amount of data available on the internet, has led to a revolution in ML research and applications. These advances have massive potential for military applications at the tactical level, particularly in improving situational awareness and speeding kill chains. One opportunity for the application of ML to an existing problem set in the military is in the analysis of Synthetic Aperture Radar (SAR) imagery. Synthetic Aperture Radar imagery is a useful tool for imagery analysts because it is capable of capturing high-resolution images at night and regardless of cloud coverage. There is, however, a limited amount of publicly available SAR data to train a machine learning model. This thesis seeks to demonstrate that transfer learning from a convolutional neural network trained on the ImageNet dataset is effective when retrained on SAR images. It then compares the performance of the neural network to shallow classifiers trained on features extracted from images passed through the neural network. This thesis shows that cross-modality transfer learning from features learned on photographs to SAR images is effective and that shallow classification techniques show improved performance over the baseline neural network in noisy conditions and as training data is reduced.
Captain, United States Marine Corps
Approved for public release. distribution is unlimited