Application of local and global unsupervised Bayesian classification algorithms to the forest

OQCompares the performances of various unsupervised Bayesian classification algorithms and their ability to distinguish different classes of trees. Two kinds of algorithms are tested: global methods, for which a Gibbs model is used to describe the class image and for which the pixels of each class are supposed independent; local methods, which only need the pixel neighborhood and can also use the correlation information between pixels. Unsupervised Bayesian classification needs two steps, one for the parameter estimation of each local or global model and one for the Bayesian classification itself. An area of the Paimpont forest in Brittany was selected, because the ground truth was available. Twenty classes have been chosen by botanists including: conifers and broad-leaved trees at different height, bare soil, copse. A comparison between these algorithms is performed.<
>