Self-Supervised Deep Learning for Radar Moving Target Indicator

This work investigates a method for self-supervised neural network training for moving target indicator radar in order to replace or augment conventional algorithms such as space time adaptive processing (STAP). Moving target indicator radar provides a critical ability to detect moving targets at long range, high area rates, and in all weather. Existing algorithms such as space-time adaptive processing work well in many cases but rely on assumptions that are often broken in practice, causing loss in performance or failure. In this work, we propose a self-supervised training scheme for neural network-based scene reconstruction and moving target detection that does not require labels. We do this by training a neural network to convert the complex-valued, multichannel radar image to produce the inputs to an auto-differentiable radar simulator (an approximate inverse solution to radar image formation). The loss function then compares the simulated and true radar images, removing the need for labels, and a simple regularizer is used to segregate clutter from targets in the simulator inputs. We compare the resulting performance to conventional algorithm on simulated data.