Complex-Valued Neural Networks for Synthetic Aperture Radar Image Classification
Theresa Scarnati, Benjamin Lewis
Abstract
Synthetic aperture radar (SAR) is an imaging modality used for a variety of military and civilian tasks, many of which could benefit greatly from computer automation. The increase in machine learning-based computer vision techniques in recent years has created a number of helpful methods to advance this goal, but most of these algorithms are designed for electro-optical RGB imagery. Applying these off-the-shelf algorithms to magnitude only SAR imagery has shown promise in tasks such as automatic target recognition (ATR). However, very few algorithms exploit the complex-valued nature of radar data. We present a survey of several complex neural network techniques as applied to a SAR data set consisting of military targets. We comment on the merits of each approach and demonstrate the accuracy of each technique when 1) training data are limited, and 2) when the training and testing data exhibit a domain mismatch.