DATA AUGMENTATION FOR SYNTHETIC APERTURE RADAR USING ALPHA BLENDING AND DEEP LAYER TRAINING

Human-based object detection in synthetic aperture RADAR (SAR) imagery is complex and technical, laboriously slow but time critical—the perfect application for machine learning (ML). Training an ML network for object detection requires very large image datasets with imbedded objects that are accurately and precisely labeled. Unfortunately, no such SAR datasets exist. Therefore, this paper proposes a method to synthesize wide field of view (FOV) SAR images by combining two existing datasets: SAMPLE, which is composed of both real and synthetic single-object chips, and MSTAR Clutter, which is composed of real wide-FOV SAR images. Synthetic objects are extracted from SAMPLE using threshold-based segmentation before being alpha-blended onto patches from MSTAR Clutter. To validate the novel synthesis method, individual object chips are created and classified using a simple convolutional neural network (CNN); testing is performed against the measured SAMPLE subset. A novel technique is also developed to investigate training activity in deep layers. The proposed data augmentation technique produces a 17% increase in the accuracy of measured SAR image classification. This improvement shows that any residual artifacts from segmentation and blending do not negatively affect ML, which is promising for future use in wide-area SAR synthesis. Outstanding Thesis Major, United States Air Force Approved for public release. Distribution is unlimited.