Synthetic aperture radar image processing by spatial-frequency representations

Synthetic aperture radar (SAR) is a microwave radar system that is carried either by air or space platforms and utilizes increased resolution for remote sensing in both civil and military areas. SAR can function in rainy or cloudy weather conditions as it uses microwave radar and can benefit from its penetration property at heavy vegetation and bare soil areas. SAR imaging can be enriched by different microwave frequencies, different electromagnetic polarizations and interferometry. Study fields such as land use classification, pollution detection, ice field monitoring, military target detection, disaster aftermath evaluation etc. can benefit from SAR imaging. SAR processing can be divided into two main categories; SAR image construction and SAR image processing. Constructed SAR image is heavily contaminated by speckle noise that is caused by multipath scattering in a resolution cell. Speckle noise can be modelled as a multiplicative noise. This work mainly focuses on comparative analysis of proposed methods on different aspects of SAR image processing steps. Speckle noise reduction methods defined in the literature are analyzed comparatively with the proposed methods. Proposed methods are based on speckle reducing anisotropic diffusion and extends the given method by means of increased noise reduction capability with Gauss filter enhancement and by means of edge preservation capability with Canny edge detection enhancement. Local Fourier transform for SAR image classification is defined and given comparatively with the standard methods. Curvelet based two feature extraction methods are adapted for SAR image classification and a third curvelet based method named histogram of curvelets is genuinely defined. Results for each method are given comparatively with standard methods. Proposed feature extraction methods are also analyzed by different clustering algorithms by means of their effect on clustering. Given the proposed methods' accuracies it can be concluded that the proposed methods are robust against speckle noise as they can be applied without speckle noise reduction and reveal better classification accuracies up to 5% higher than speckle reduced data accuracies. Proposed methods are in the form of spatial and local feature extractors based on their definitions, the results also emphasize the importance of spatial local features.