A COMPARISON OF DATA-MINING TECHNIQUES INPREDICTIVE SOIL MAPPING

摘要

To predict soil maps, data-mining techniques can be utilised. The aim of these techniques is to extract hidden predictive knowledge from large databases. In terms of soil science they are able to learn the relationship between mapped soil classes as well as soil-forming factors, which can be used to predict soil classes in comparable landscape units. Thus, it is possible to automatically build reproducible digital soil maps, helping to speed up field mapping and to reduce costs. The main objective of this chapter is to compare the ability of different data-mining techniques and algorithms from statistics and information theory, including artificial neural networks (ANNs), support vector machines (SVMs), linear regression, learning vector quantisation and classification trees. The techniques are discussed in terms of prediction accuracy and usability for GIS-based usage. Altogether 10 data-mining algorithms were tested to predict soil classes on the basis of 65-terrain attributes. Prediction accuracy is tested inside and outside the learning area to compare their generalisation ability