Recent Advances in Synthetic Aperture Radar Enhancement and Information Extraction

Abstract

Synthetic Aperture Radar (SAR) systems are all-weather, night and day, imaging systems. Automatic interpretation of information in SAR images is very difficult because SAR images are affected by a noise-like characteristic called speckle that arises from an imaging device and strongly data and makes automatic image interpretation very difficult. The speckle noise in SAR images can be removed using an image restoration technique called despeckling. The goal of despeckling is to remove speckle-noise from SAR images and to preserve all image’s textural features. The statistical modeling of SAR images has been intensively investigated over recent years. In statistical image processing an image can be viewed as the realization of a joint probability density function. Since joint probability functions have analytical forms and few unknown parameters usually, the efficiency of the denoising algorithm depends on how well the chosen model approximates real data. The wavelet Daubechies (1992) based despeckling algorithms are proposed in Dai et al. (2004), Argenti et al. (2006), Foucher et al. (2001). The second-generation wavelets like Contourlet Chuna et al. (2006) have appeared over the past few years. Despeckling using Contourlet transform Li et al. (2006) and Bandelet Sveinsson & Benediktsson (2007) transforms show superior despeckling results for SAR images compared with the wavelet based methods. Model based despeckling mainly depends on the chosen models. Bayesian methods have been commonly used as denoising methods, where the prior, posterior and evidence probability density functions are modeled. The image and noise models in the wavelet domain are welldefined using the results in Argenti et al. (2006), Gleich & Datcu (2007) and the noise free image is estimated using a MAP estimate. The speckle noise in the SAR images is considered as a multiplicative noise Walessa & Datcu (2000), and can be also presented as a signaldependent additive noise Argenti et al. (2006). The log transformed image is modeled using zero location Cauchy and zero-mean Gaussian distributions in order to develop minimum means absolute error estimator, and maximum a posteriori estimator. This paper presents the state-of the art methods for information extraction and their comparison in efficiency of despeckling and information extraction. This paper presents three methods for despeckling and information extraction. The first method is wavelet-based despeckling and information extraction method using the General Gauss-Markov Random Field (GGMRF) and Bayesian inference of first and second order. The second and third methods use the GMRF and Autobinomial model with the Bayesian inference of first and second order. The despeckling performance is compared and the texture parameters estimation is presented.