Applications of simulated annealing to SAR image clustering and classification problems

Abstract

This paper deals with simulated annealing applications to some unsupervised classification problems. First, we present an adaptation of the simulated annealing technique to the clustering problem, and compare its results with those provided by classical c-mean clustering algorithms, which lead only to local optimality. It is shown that simulated annealing technique yields improved results over the standard c-means method.

Then, we compare the classified images obtained from two different clustering algorithms: simulated annealing, and K-means guided by initialization from optical data, applied to MAC-Europe'91 images. The study reveals a poor robustness of SAR image classification versus the estimated duster characteristics, and shows that most of culture types were best identified by the global optimization.