Machine learning in soil classification

In a number of engineering problems, e.g. in geotechnics, petroleum engineering, etc., intervals of measured series data (signals) are to be attributed a class maintaining the constraint of contiguity and standard classification methods could be inadequate. Classification in this case needs involvement of an expert who observes the magnitude and trends of the signals in addition to any a priori information that might be available. In this paper an approach for automating this classification procedure is presented. Firstly, a segmentation algorithm is applied to segment the measured signals. Secondly, the salient features of these segments are extracted using boundary energy method. Based on the measured data and extracted features classifiers to assign classes to the segments are built; they employ decision trees, ANNs and support vector machines. The methodology was tested for classifying subsurface soil using measured data from cone penetration testing and satisfactory results were obtained.