Classification of multi-temporal SAR images and Insar coherence images using adaptive neighborhood model and simulated annealing approach

Abstract

A classification algorithm for multi-temporal SAR images and InSAR coherence images is developed. This is done by minimizing an energy function which is defined by the sum of (1) the difference between the pixel value and the cluster center (mean value of a class) and (2) the difference between the pixel label (class) and its neighborhood. The global minimization of the energy function of the whole image is achieved by a simulated annealing approach. The cluster centers for different classes are calculated using fuzzy c-mean and possibility c-mean algorithms or determined manually from the image values. In order to preserve the details of the original images, a neighborhood structure with a 5*1 window is used to calculate the energy function, whose orientation is determined by the local homogeneity (e.g. the variance of the pixel values within the window) of the original images. This algorithm is used to classify three ERS-2 SAR images over the Mekong delta region, Vietnam. By comparing visually the color composite SAR images and the classified image, it is shown that the classification result is reasonable. Moreover, this algorithm is also used to classify an InSAR coherence image acquired from two JERS SAR images over south Sumatra, Indonesia. The classified InSAR coherence image delineates the bare land and forest clearly, since the coherence of bare land is larger than that of forest.