Change Detection for Urban Areas in High Resolution Sar Iamges Using Second Kind Statistics Based G0 Distribution

It is very popular to monitor the changes of urban areas via synthetic aperture radar data, which is rarely influenced by weather conditions or imaging time. However, change detection using multi-temporal SAR data may confront with difficulties, due to the highly speckled noise generated by the coherent imaging mechanism. Thus SAR images change detection algorithms mainly contains three steps: (i)modeling the SAR image RCS as well as clutters by some proper statistical methods; (ii)generating a change map by computing a certain similarity measure based on the former model; (iii)masking the change map via proper thresholding techniques to obtain the final change areas. Quite a few researches have been carried out for SAR images change detection. Considering the multiplicative speckle noise of SAR images, the most common change detection algorithm is the ratio or log ratio method [1].The Kullback-Leibler divergence based algorithm [2], using the Pearson Distribution Family to approximate the SAR clutter within the neighborhood of each pixel in the image, distinguishes a volcano influenced areas between two 10 m resolution Radarsat images acquired before and after the eruption of the volcano. A generalized Gaussian model based change detection algorithm is performed after a noise reduction preprocessing step between chips from two multi-temporal ERS2 images of a city area via a reformed version of Kittler-Illingworth threshold selection criterion [3]. Before proposing a change detection method, one has to carefully examine the statistical models and choose the most appropriate one for the problem at hand. The well known Gaussian model [4] of fully developed speckle and many models derived from it are on the basis of a large amount of random reflectors per resolution cell hypothesis. However, with the increasing of the image resolution, the RCS and clutter of urban areas may deviate from the Gaussian model, especially when the area is man made or the resolution cell is about the size of the objects. Many non-Gaussian models(K, Weibull, Log-normal, Nakagami-Rice etc) are proposed to fit with the scattering statistics, but none of them is flexible enough to model the surface of urban