Case Studies with Sar Data for Assessing the Utility of Manual Feature Selection in Machine Learning

An important question in the application of machine learning models to SAR ATR is the degree to which data is processed prior to training and testing. This work examines performance at two classification tasks when data features are manually-selected as model inputs and compares the performance at the same task when unprocessed, complex-valued pixel data is used as the input. First, the ability of a convolutional neural net to classify subaperture behavior in a SAR image was compared when spatial versus frequency data was given as the input. Second, the effect of super-resolving point targets in SAR imagery was investigated when used as input to a convolutional neural net trained to identify the number of targets in a scene. Using manually-selected features in the subaperture experiment show no benefit in classification accuracy, but it did show improvement in learning rate. However, super-resolving point targets showed a significant accuracy benefit.