Parameter Selection in Non-quadratic Regularization-based SAR Imaging
Sabanci University Research Database
Many remote sensing applications such as weather forecasting and automatic target recognition (ATR) require high-resolution images. Synthetic Aperture Radar (SAR) has become an important imaging technology for these remote sensing tasks through its all-weather, day and night imaging capability. However the effectiveness of SAR imaging for a speciﬁc decision making task depends on the quality of certain features in the formed imagery. For example, in order to be able to successively use a SAR image in an ATR system, the SAR image should exhibit features of the objects in the scene that are relevant for ATR. Recently, advanced SAR image formation techniques have been developed to produce feature-enhanced SAR images.
In this thesis, we focus on one such technique, in particular a non-quadratic regularization-based approach which aims to produce so-called “point-enhanced SAR images”. The idea behind this approach is to emphasize appropriate features by means of regularizing the solution. The stability of the solution is ensured through a scalar parameter, called the regularization parameter, balancing the contribution of the data and the a priori constraints on the formed image. Automatic selection of the regularization parameter is an important issue since SAR images are ideally aimed to be used in fully automated systems. However this issue has not been addressed in previous work.
To address the parameter selection problem in this image formation algorithm, we propose the use of Stein’s unbiased risk estimation, generalized cross-validation, and L-curve techniques which have been mostly used in quadratic regularization methods previously. We have adapted these methods to the SAR imaging framework, and have developed a number of numerical tools to enable their usage. We demonstrate the effectiveness of the applied methods through experiments based on both synthetic as well as electromagnetically simulated realistic data.