RF Modeling for Weapon Systems ML-ATR Algorithm Training

The main objective of this effort is to develop tools and techniques to augment synthetic data generated from Xpatch at Ku- and Ka-bands to train a machine learning automatic target recognition (ML-ATR) algorithm that performs well on highly diverse data sets ľ not just on the training data set. The augmentation methods developed under the effort will transform pristine Xpatch signatures to appear more like measured data from the perspective of an ML-ATR algorithm. The data augmentation techniques will address shortcomings such as known to over-predictions of RCS for complex structures. The augmentation techniques will also induce quadratic phase error that is present in measured SAR imagery as well as impart defocusing on the SAR images to generate more ôreal-worldö artifacts in the data. We will develop these data augmentation techniques combining custom-developed and COTS tools. Once the augmentation tools are developed and proven, we will render Xpatch synthetic signatures of up to 10 ground mobile and 10 maritime targets and then apply the augmentation techniques to produce much more realistic and less pristine synthetic data ľ data that can be used to train and test an ML-ATR algorithm. Finally, we will use the synthetic data generated with Xpatch and augmented with the tools developed under this effort to train and test a high-performing SAR radar ML-ATR algorithm. We will use measured data to train the network and then test the network with both the pristine Xpatch synthetic data and the augmented synthetic data, computing the Probability of Identification (PID) for each. Higher PID for the augmented data (compared to the Xpatch pristine data) would demonstrate the success of the effort. The final effort will be to train the same network using only the augmented Xpatch data, and then test the networkĺs performance using the measured data.