Remote Sensing Image Change Detection Algorithms Based On Multi-Objective Optimization

The detection of change is one of the most important tasks in remotesensing analysis, which refers to extracting change regions in a pair ofimages acquired on the same scene at different time. Change detectiontechnologies have been proved particularly useful in environmentalmonitoring, agricultural surveys, urban studies, forest monitoring and soon. There are two novel change detection methods proposed in this paper:(1) An unsupervised change detection approach by minimizing themean square error is proposed. In single-band images, the accurate solutionof the change mask with minimum mean square error can be obtained in anacceptable time. In multispectral images, it is considered as amulti-objective optimizations problem. Genetic algorithm is used to obtainthe optimal compromised solution. The simulation experiment is designedto analyze the method. The total error rate and Kappa coefficient of thechange detection result shows the proposed method performs quite well. Italso shows accurate radiometric correction of images is required.(2) An unsupervised change detection approach based oncross-correlation coefficient is proposed. The new method is robust toradiometric difference. According to the binary change detection mask, thechange detection problem can be modeled as partitioning two input imagesinto two distinct regions, namely “changed” and “unchanged”. Each regionin the pair of the images is considered as two sets of variables, whosecross-correlation coefficient is calculated in order to provide an optimalpartition of the changed and unchanged regions. In the optimal partition,the cross-correlation coefficient of the set of the unchanged variables isclose to1, while that of the changed variables is close to0.. The optimalpartition is a multi-objective optimizations problem, a genetic algorithm isused to obtain the final optimal compromised solution. The iterativealgorithm is proposed to overcome the problem in change detection basedon cross-correlation coefficient that the set of variables whosecross-correlation coefficient is close to1does not always represent the unchanged region. It is because sometimes the value of pixels is changedfrom the intensity level into another intensity level. In each iterative step,the remained regions are also partitioned into two distinct regions. The setsof variables whose cross-correlation coefficient is close to1are extractedand judged whether they belong to the unchanged region. The simulationexperiment shows that the result using the new method is effective androbust to radiometric difference.Finally, two proposed methods are applied to real remote sensingimages. Two input images are acquired by the Thematic Mapper of theLandsat5satellite on the island of Sardinia, Italy, in September1995andJuly1996, respectively. The reference change error matrixes show theproposed methods are effective and accuracy.