Effect of Spatially Variant Apodization on SAR Image Classification

Abstract

This paper investigates the effect of the application of Spatially Variant Apodization techniques to SAR images on the statistical properties of the apodized radar signal and on the capability of extracting information from apodized SAR images. After deriving the statistical model of the apodized image (in terms of both probability density function and moments) two new classification schemes of homogeneous regions with different radar cross section in apodized SAR images are obtained. The performance of the new schemes are deeply investigated and compared with the performance achievable by Maximum Likelihood classification schemes developed under the assumption of Gaussian statistics and applied to the original images and to the apodized images. The performance analysis shows that the new schemes maintain the information extraction capabilities while at the same time allowing the sidelobe level to be reduced and the mainlobe resolution to be preserved. Radar imaging often requires sidelobe control: as well known both linear and non linear techniques can be used to reduce the sidelobe level. Linear techniques are based on the use of amplitude weighting function (frequency domain) before the final Fourier transform: in this case the reduction in sidelobe level is obtained at the expense of the mainlobe width with a loss in resolution. Apodization techniques are non linear techniques which have been proposed to reduce sidelobe level while preserving mainlobe resolution. This is of particular importance especially in range dimension where the high resolution is provided by the transmitted bandwidth, usually limited by technological or regulations constraints. Due to their non linear behavior, it is of great importance to understand the impact of the apodization techniques on the quality of SAR (Synthetic Aperture Radar) images in terms of capability of extracting information from apodized SAR images. In this paper we focus on Spatially Variant Apodization (SVA), (1), applied separately to the in-phase (I) and quadrature (Q) components in the range dimension and analyze the impact of SVA on SAR image statistical properties and on information extraction in terms of classification of homogeneous regions with different radar cross section. probability density function (PDF) of the apodized random variable (r.v.) and (ii) the moments of generic order of the apodized r.v.. On this basis the impact of apodization techniques on the following interpretation of SAR images is analyzed in Section 3. In particular we aim at understanding the impact of SVA on the classification of homogeneous regions in SAR images. We show that the use of classical ML (Maximum Likelihood) classifier, developed under the assumption of a Gaussian statistics, results in strong losses when applied to apodized SAR images; to cope with this problem, by using the knowledge of the statistical properties of the apodized SAR images, we derive two new classification schemes in order to maintain the information extraction capabilities while at the same time allowing the sidelobe level to be reduced and the mainlobe resolution to be preserved. Finally, some conclusions are drawn in Section 4.