Random forest classifiers for hyperspectral data

Two random forest (RF) approaches are explored; the RF-BHC (binary hierarchical classifier) and the RF-CART (classification and regression tree). Both methods are based on a collection (forest) of tree-like classifier systems where the difference is in the way the trees are grown. The BHC approach depends on class separability measures and the Fisher projection, which maximizes the Fisher discriminant where each tree is a class hierarchy, and the number of leaves is the same as the number of classes. The CART approach is based on CART-like trees where trees are grown to minimize an impurity measure. Here, these different RF approaches are compared in experiments. The RF approaches were investigated in experiments by classification of an urban area from Pavia, Italy using hyperspectral ROSIS (reflective optics system imaging spectrometer) data provided by DLR.