Unsupervised Bayesian classification of remote sensing images using a pyramidal structure and the SEM algorithm

The pyramidal image decomposition consists of a succession of filtering and undersampling giving little smoothed images. Details are removed allowing a better classification at each level of the pyramid. The SEM, a stochastic version of the EM algorithm, has been chosen for that. At some level, pixels are split into two categories: those, which come from parents already classified and those, which have not been previously classified. The classification is done on these last. Then, according to an homogeneous test, some of these pixels are considered as well classified and the other will be classified at the next level again. Going from the top to the bottom of the pyramid leads to the whole image classification. This procedure has been studied for various conditions of filter type, homogeneous criterion and applied to simulated and remote sensing images