Remote sensing the seabed has become an indispensable issue for marine environmental studies. Acoustic technologies are the more efficient and cost effective devices to achieve this task. Among these technologies, single-beam echo-sounders (SBES) are widely used because they are inexpensive, practical and non invasive. Moreover, the data recorded by SBES can be calibrated and may be used to automatically classify and to characterize the seabed. Many studies exist on this subject and first, authors point out that existing seabed classification methods from SBES data suffer from misclassification due to artefacts. Indeed, the recorded acoustic signal not only depends on seabed types but also on characteristics of the acquisition system, the propagation medium and the morphology (depth, slope) of the seafloor. Then, several authors have concluded that standard approaches of seabed classification are not optimal. According to a recent paper, the cause of the problem is that the approaches lack physical bases. Most approaches use discriminating features synthesized with statistical data analysis methods such as principal component analysis and thus with no link with the physical phenomenon. Thus, in this paper, we investigate a methodology of seabed classification based on the physical properties of the backscattered acoustic signal. A physical model is used to generate echoes for various types of seabed, in different configuration of depth and slope. A study is carried out to assess the effect of the pre-processing steps (echoes averaging, depth and slope) on echoes. Then, features are extracted from echoes based on previous studies and new features are proposed. Finally, from the linear discriminant analysis, a linear combination of features is found independent of the slope and the depth.