A decision tree approach has been developed that gives us a better understanding of the predictive structure of two fundamentally different problems: surname pronunciation generation and scenic beauty estimation of forestry images. Decision trees have been constructed using a Bayesian approach which incorporates conditional information to maximize the posterior probability. For the surname problem, we achieved an error rate of 39%, compared to 40% previously best reported result achieved by a recurrent neural network approach. For the scenic problem, we achieved an error rate of 43%, compared to 36% for a Principle Components Analysis approach. In the latter case, decision tree performance is shown to suffer due to features that have little descriptive power. Clearly, more robust features are required to improve the performance of the decision tree classifier.