The work described in this paper is situated in the field of data processing and pattern recognition applied to impulse radar and ultrasonic signals. The subject of the study, presented here, is the identification of the type of underground medium (mine, sand, stone,...) under the recording sensor on the basis of the signal reflected to it. A data pre-processing filter ensures noise reduction. The filtered data are transformed into a vector of features. For the design of a classification method, this feature vector is reduced so that it only contains those features that really contribute to the discrimination between the different classes. This is done in a “learning phase”, using measured signals with known classification. Finally, the classification method is applied to the calculated reduced feature vectors derived from measurement signals with a priori unknown class membership. Section 2 briefly describes the pre-processing filter, sections 3 and 4 explain the feature selection and classification methods. Finally, section 5 presents a few experimental results, which are further discussed in section 6.