Analysis of NNs Detectors for Targets with Unknown Correlation in Gaussian Interference

This paper carries out a study of the suitability of Neural Networks (NN) as solutions for the problem of detecting Gaussian targets with unknown one-lag correlation coefficient (ρ_s) in different radar clutter environments (Additive White Gaussian Noise (AWGN) and correlated Gaussian clutter plus AWGN). The optimum Neyman-Pearson detector is formulated assuming an uniform variation of ρ_s ∈ [0; 1]. As this solution conveys a complex integral, sub-optimum approaches based on the Constrained Generalized Likelihood Ratio, CGLR, are proposed as reference ones. Detectors based on different NN arquitectures have been designed, using supervised techniques and target patterns with ρ_s varying uniformly in [0, 1]. Results prove that among those considered, Second Order NNs present a great robustness for the considered cases of study, although other NN architectures can be more suitable for specific cases of study, so they are less robust but present better detection performance and/or lower computational cost.