To derive efficient predictions from a learning set provided by a geographical information system (GIS) needs to take into account errors that occur because of the poor quality of source maps and error propagation through GIS procedures. Using classification tree analysis, our objective was to build an approach which could take these errors into account, and so provide robust prediction rules for mapping natural systems. Our application addresses soil unit predictions from topographical and geological data. The learning set is taken from a sample area located in a French Mediterranean valley in which a large scale soil survey was conducted. We propose a rule for only conserving, among the splits of an initially large classification tree, the ones that provide a significant gain of purity (i.e. precision of soil predictions). This rule needs an estimate of the standard error on the purity index. This is achieved by evaluating the errors in the learning set, and then studying their propagation through the classification tree. The severity of the rule can be adjusted according to a risk threshold defined by the user. Our results indicate that the new rule has a strong influence on classification tree size. Furthermore, comparison with actual soil maps indicates that this rule provides more robust predictions than a classical pruning technique used in classification tree analysis. The soil predictions we obtain provide a medium scale soil map.