This paper deals with the unsupervised segmentation of textured seafloors from sidescan sonar imagery. The classical approaches for texture analysis are based on the estimation of co-occurrence matrices (which express the distribution of co-occurring values at a given offset) and the Haralick features derived from. However, the GLCM is strongly dependant on the parameterization of the offset (e.g. the distance d and the angular direction θ for the estimation of the number of co-occurring values). In most cases it is not at all obvious how such a choice should be made for (d, θ) and it is even more difficult to arrange for it to be made automatically. In this paper, we investigate the ability of another approach based on spectral features to discriminate between seabed textures. Spectral features are estimated from directional filter bank (DFB) in the 2DFourier space. A subsequent analysis of the pattern isotropy is conducted by dividing the medium spectral band into small, overlapped, angular sectors. The unsupervised segmentation used is a modified version of Kohonen SOFM (Self-Organizing Feature Maps) with splitting process of images to take into account the grazing angle dependency. The data used in our tests are sonar images recorded by Klein 5000 sidescan sonar.