The Random forest approach for land cover mapping

In this paper, we proposed a new random forest algorithm designed specifically for the land cover mapping problem. Three approaches are investigated, namely, pixel-based, neighbor-looking and combination of both. In the pixel-based approach, we use the fact that all decision trees are different whereas, in the neighbor-looking, the decisions from neighboring pixels are used when the decisions from the Random forest is not clear. Our results on simulated and actual data set showed that our new RF approaches outperformed the traditional one.