Affinity Propagation for Unsupervised Classification of Remotely Sensed Images

The aim of this paper is to present a new unsupervised classification method for satellite multispectral images based on affinity propagation (AP) algorithm. Recently proposed, affinity propagation becomes the most widely methods for data clustering. This technique is essentially based on passing of messages between pixels to be automatically classified without any a priori knowledge about the number of classes. Its main advantage is that initially all pixels to classify are considered as centroids or "exemplars". However, the AP method has two major drawbacks: 1) when it comes to partition multispectral images of high spatial size, complexity of computation becomes quadratic 2) it gives an overestimation of class number due to its great sensitivity to very small variations in the image. In this work, we present the AP algorithm in its original version and we have proposed an iterative AP-block procedure to address the two issues mentioned above. Both versions have been applied to classify a low spatial resolution image acquired by ETM+ sensor of american satellite LandSat-7. From obtained results, it is concluded that the proposed AP-block classifier is more appropriate and more efficient to unsupervised image classification than the classical AP algorithm.