Abstract:

In this paper, we propose a novel method for unsupervised change detection in multi-temporal satellite images by using multiobjective cost function optimization via genetic algorithm (GA). The spatial image grid of the input multi-temporal satellite images is divided into two distinct regions, representing ``changed'' and ``unchanged'' regions between input images, via the intermediate change detection mask produced by the GA. The dissimilarity of pixels of ``changed'' regions and similarity of pixels of ``unchanged'' regions between input multi-temporal images are measured using image quality metrics which consider correlation, spectral distortion, radiometric distortion, and contrast distortion. The contextual information of each pixel in intermediate change detection mask is considered by employing binary information around its 3 \times 3 neighborhood. A multiobjective cost function for the intermediate change detection mask is obtained by considering contextual information, similarity and dissimilarity measures. The final change detection mask is achieved through minimization of the multiobjective cost function via different realizations of intermediate change detection masks produced by the GA. The proposed method does not need to compute difference image from multi-temporal satellite images, thus can be used as a general purpose change detection method for both synthetic aperture radar (SAR) and optical images. Change detection results and comparisons with the state-of-the-art techniques are shown on multi-temporal georeferenced SAR images acquired by ESA ERS-2 on the city of San Francisco, California, and optical images acquired by Landsat 5 TM on part of Alaska.

Keywords: