POSTER: Analysis of soil data using Classification and Regression Trees

Analysis of soil data using Classification and Regression Trees

S. Caetano, J. Smeyers-Verbeke, Y. Vander Heyden*

FABI, Department of Analytical Chemistry and Pharmaceutical Technology, Vrije Universiteit Brussel- VUB, Laarbeeklaan 103, 1090 Brussels, Belgium
*e-mail: yvanvdh@vub.ac.be, tel.: +32 2 477 473 4, fax: +32 2 477 47 35

This poster illustrates the analysis of soil data, provided by WP1 (partner 23), done with Classification and Regression Trees (CART).
The aim of this work was not only to determine whether it was possible to classify soil samples based on their concentrations of trace elements and isotopic ratios, but also to verify that CART was a good chemometrical technique to reach that goal.
CART is a non-parametric procedure, i.e., it makes no assumption about the distribution of the data, and was introduced for explaining and/or predicting both categorical and continuous responses.
The models obtained for lithology (sandstone, limestone and shale), for both top and subsoil show that it is possible to classify the samples just by using two variables (Aluminium and Tantalum 181, in the case of topsoil, and Silicon and Tantalum 181, for subsoil). Both classification trees give quite good results, showing probabilities that a new sample is misclassified of 2.5% and 4.8% for top and subsoil, respectively.
Classification according to sampling site (sites from 1 to 9) for the two types of soil is also possible. In these cases eight variables are needed to classify the samples.
When classifying the sampling teams (team 1 to 4), the models gave worst results, showing that, based on trace elements or isotopic ratios, the sampling teams are not really distinguishable.
Also the content of organic matter was modelled with acceptable results. This demonstrates CART's capacity to model continuous responses.