Aspect segmentation and feature selection of radar targets based on average probability of error

Through statistical estimation by a non-parametric model, a fused polarimetric and resonant return from the radar target is modelled as a function of the target aspect angle. The outcome of this type of modelling is a set of non-parametric density estimates, which are then used to represent this target in a multi-dimensional probability space. These densities within this probability space can be well separated and therefore utilised to make decision rules to identify targets of interest. The return set to be modelled is the average power set associated with spectral bands centred on the target natural resonant frequencies. This return set is mapped into density set using a Gaussian kernel function; subsequently, the density set will be considered as the target radar feature set of interest. To decrease density overlapping between respective densities of different targets, a criterion based on the Bayesian error is employed; first, to bisect the aspect global range into smaller sectors, and second, to select discriminative features that can minimise the average probability of error between the targets respective features. The results show that two targets with similar resonant frequencies can be separated by the Bayesian error criterion based on the proposed features. A simple likelihood ratio test had more than 80% success down to 20%dB of signal-to-noise ratio.