A comparison of SAR image speckle filters

High quality images of Earth produced by synthetic aperture radar (SAR) systems have become increasingly available, however, SAR images are difficult to interpret. Speckle reduction remains one of the major issues in SAR imaging process, although speckle has been extensively studied for decades. Many reconstruction filters have been proposed and they can be classified into two categories: multilook and/or minimum mean-square error (MMSE) despeckling using the speckle model; and maximum a posteriori (MAP) or maximum likihood (ML) despeckling using the product model. The most well known Lee, Kuan, and Frost filters belong to first category. These filters are based on conventional techniques that were originally derived for stationary signals, such as MMSE. In the second category, filters are based on the product model, such as the MAP Gaussian filter and the Gamma filter, and require knowledge of the a priori probability density function. These filters force speckle to have nonstationary Gaussian or gamma distributed intensity mean. The speckle filtering is mainly Bayesian model fitting that optimizes the MAP criteria. Scene reconstruction is performed using an inversion of the ascending chain. An objective measure is required to compare the technical merits of these filters, and Shi et al. presented a comparison 15 years ago. In this paper, a brief introduction of speckle, product, and filter models is summarized. A review of some most widely used SAR image speckle filters is given. And stationary speckle filters, like Lee, Kuan, and Frost filters, and nonstationary speckle filters like Gamma MAP filter are studied. Despeckling results on stationary and nonstationary SAR image of these speckle filters are presented.