This paper addresses the problem of target detection in the presence of Doppler spread clutter which is neither stationary during a coherent integration time (CIT) interval nor across different range bins. This phenomenology occurs, for example, in over-the-horizon skywave HF radar where propagation through moving ionospheric inhomogeneities spreads the surface clutter and the clutter statistics can change quite abruptly during a CIT and across range bins. In these cases, the performance of conventional adaptive techniques suffers from a lack of adequate training data. The method proposed here breaks the full CIT into smaller sub-CIT's which are then extrapolated using low order AR models. The Doppler spread clutter is thus effectively modeled as an abruptly time varying autoregressive (ATVAR) process. Subsequent Doppler processing and coherent combining of the extrapolated sub-CIT's is then performed with improved signal-to-clutter gain since only a small proportion of the sub-CIT's are corrupted by abrupt non-stationary behavior. Moreover, nearly full coherent signal gain against noise is maintained. Initial processing on experimental radar clutter data with injection of a simulated target illustrates that this approach can provide an SCNR improvement of more than 5 dB compared to conventional Doppler processing. © 2006 IEEE.