Combining non-parametric models for multisource predictive forest mapping

Most models of forest type for predictive mapping cannot produce estimates of confidence in the prediction of individual pixels, even where they provide good overall accuracy. A new strategy that combines several models based on different principles not only provides estimates of prediction confidence, but also improves the mapping accuracy. In this study, the theoretical foundation of Artificial Neural Networks, Decision Trees, and Dempster-Shafer's Evidence Theory are briefly reviewed, compared, and applied to a common data set. Two ways for integrating the results of the three models were then evaluated. One method was to separately harden the probability results of the three models, then combine them to make a single classification. In the second method, the probabilities of the three models for each pixel were simply averaged, then hardened to a single classification. Deferring the hardening to the final stage produced the best results. The 3 percent increase in overall accuracy for the second approach compared with the best individual model is encouraging. More importantly, estimates of prediction confidence were derived, based on a comparison between a combined model and the three models, something that is impossible using a single model.