Multiscale Feature Deep Fusion Method for Sea-Land Clutter Classification

Sea-land clutter classification (SLCC) is a key technology to improve the target positioning accuracy of skywave over-the-horizon-radar (OTHR). The scale of the sea and land regions in the range-Doppler (RD) map changes dynamically as the OTHR detection area changes. The existing SLCC methods are only applicable to the scenario of sea-land clutter from a single azimuth-range cell and do not consider the multiscale characteristics of the RD map, resulting in poor classification performance. To solve the above problems, this article proposes an SLCC method based on multiscale feature deep fusion, namely, an attention-aided pyramid scene parsing network (AAPSPNet). First, an attention mechanism is adopted to make AAPSPNet effectively learn the features near the 0-Hz frequency of the RD map. Second, the image pyramid is used to fuse the multiscale features, enabling effective utilization of context and global information in the RD map. In the data preprocessing, a narrowband radio frequency interference (NRFI) elimination method based on information entropy (NRFIIE) is proposed. The NRFI in the RD map can be eliminated and the classification accuracy of AAPSPNet can be improved. To verify the effectiveness of AAPSPNet, we build a sea-land clutter original dataset, a sea-land clutter scarce dataset, and an NRFI dataset. Compared with state-of-the-art SLCC methods, the proposed AAPSPNet has the best performance on the given datasets. Meanwhile, the effectiveness of NRFIIE is verified on the NRFI dataset.