Combining Non-Parametric based models for multisource predictive forest mapping

Using a single model for forest type predictive mapping will not produce good estimates of confidence in the prediction of individual pixels, even with good overall accuracy. A new strategy which combines several models based on different philosophy could not only reduce the uncertainty of predictive modelling, but also improve the mapping accuracy. In our study, a Artificial Neural Networks, a Decision Trees, and a model of Dempster-Shafer’s Evidence Theory were individually applied to a common data set. Two ways for combining the results of the three models were then evaluated. One approach was to separately harden the probability results of the three models at first, then the three thematic maps were combined to a single classification map. On the second approach, before hardening, the probabilities of the three models for each pixel were simply averaged, then hardened to a single classification map. The 3% increase of overall accuracy for the second approach compared with the best individual model is encouraging. More important, a confidence map was produced based on the comparison between the three models, something which is impossible by using a single model. It was also demonstrated that deferring the hardening process to the last stage gives the greatest benefit for the combining process.