Ultra-wideband synthetic aperture radar (UWB SAR) is a sufficient approach to detect landmines over large areas from a safe standoff distance. Feature extraction is the key step of landmine detection processing. On one hand, the feature vector should contain more scattering characteristics to discriminate landmines from clutters; on the other hand, the dimension of feature vector should be lower to avoid the "curse of dimensionality". In this paper, a novel feature vector extraction method is proposed. We first obtain the scattering information in the four-dimensional domain, i.e., range, azimuth, frequency and aspect-angle, via the space-wavenumber distribution (SWD). Since the data after SWD are with higher dimension and local nonlinear structures, a typical manifold learning method, Isomap, is used to reduce the dimension. The validity of the proposed method is proved by using the real data collected by an airship-borne UWB SAR system.