STAP-informed Neural Network for Radar Moving Target Indicator

As machine learning (ML) approaches continue to advance the state-of-the-art for radar clutter removal, their network architectures leave valuable domain understanding unex-ploited. Further, there remains a dearth of sufficiently large radar datasets for practitioners to train models, compare algorithms, and report progress. We present a neural network architecture that leverages knowledge of radar image formation via locally stenciled IQ input data for rapid clutter removal in Ground Moving Target Indicator problems. The neural network replaces Space Time Adaptive Processing (STAP) and produces a clutter removal weight matrix with identical function. We also present a training and validation radar dataset of sufficient complexity to stress-test STAP-based techniques with novel ML approaches. The dataset consists of multichannel, complex IQ data under three clutter conditions, each having over 2000 range-Doppler chips for training and evaluation. It includes variable numbers of injected moving point targets with a range of scattering cross sections obfuscated by a heterogeneous clutter ridge. Variants of the dataset include a representation of windblown trees and a low signal-to-noise ratio (ground-clutter-free) dataset to directly measure and contrast noise and clutter removal performance. In general, our physics-integrated, convolutional neural network architecture outperforms STAP on all simulated datasets by a significant margin.