Cost-efficient training strategies for space-time adaptive processing algorithms

Space-time adaptive processing (STAP) usually requires the estimation of large-dimension clutter covariance matrices. The mean loss in output SNR is a function of the number of statistically similar data samples used to estimate the covariance matrix. This number is generally 3 times the dimension of the covariance matrix or more. In nonhomogeneous clutter environments it is difficult to obtain this many statistically similar data samples using a data selection rule that is computationally simple. We present several new training strategies that select data samples from as close to the target range-gate as possible and simultaneously maintain a low computation count. A "training strategy" is the rule used to select data samples for covariance matrix estimation. A new training strategy is presented along with a recursion for efficient estimation of the clutter covariance matrix at each target range-gate. Also, a new training concept called freeze training is presented and shown to reduce the number of computations and to mitigate clutter discretes in nulled output data. A computation-count comparison is presented with each training strategy.