As is commonly known, the optimal detector for any waveform in Gaussian background noise is a matched filter. However, HF atmospheric noise is non-Gaussian, necessitating alternate detector designs. The industry standard CCIR 322 model for HF atmospheric noise is a graphical, empirical model based on observations of HF atmospheric noise taken over the course of many years at numerous worldwide receive sites. In this work, it is shown that the CCIR 322 noise model may be approximated by a random process which is a member of the class of non-Gaussian random processes known as spherically-invariant random processes (SIRPs). This analytical, empirical SIRP representation is then shown to be identical to the Hall model of impulsive phenomena. In a departure from the optimal, parametric, coherent detector derived by Hall, a locally optimal, parametric, non-coherent detector is presented. In addition, a means to estimate the parameters of the Hall model is provided and is used as the basis for an adaptive, locally optimum, parametric, non- coherent detector design. Monte Carlo simulations are performed to evaluate detector performance, and comparisons are made with two common, sub-optimal, non-parametric detectors.