There are many different autofocus algorithms for correcting phase errors in stripmap SAR images. In this paper, we develop a mathematically principled model-based phase error estimation method for stripmap SAR images. The proposed phase estimation method uses classical subspace fitting techniques that are well known in the array processing literature. The novelty in this paper is the development of the proposed autofocus method and how the phase estimates are found using convex optimization. It is shown that the proposed method is non-iterative in the sense that iterations between the image domain and the range compressed domain to obtain our phase error estimates are not needed. It is also shown that the phase error estimates can be applied while simultaneously forming the SAR image using the convolution back-projection algorithm.
I. BACKGROUNDDespite the best efforts to have a synthetic aperture radar (SAR) sensor follow a predetermined nominal flight path, phase errors still exist in the SAR data that corrupt the focused image. It is well known that these phase errors come mainly from two sources, [1], [2], [3]. A low frequency phase error will exist for uncompensated platform deviation, which has the effect of broadening the main-lobe of the azimuth compressed signals. The other source of phase error stems from signal propagation effects. These phase errors tend to be high frequency phase errors and the effect is not so much a broadening of the main lobe of the azimuth compressed pulse as raising the side-lobes, which lowers the contrast of the image and masks less reflective objects. Autofocus algorithms are data driven algorithms that estimate and correct these phase errors.There are several autofocus algorithms for stripmap SAR that can be classified as parametric, non-parametric, or metric based. The parametric approaches tend to model the phase error as a polynomial or as sinusoidal and use the data to estimate the coefficients, [2], [3], [4], [5]. Since the data is used to estimate the coefficients, these methods are limited to only reliably estimating a small number of coefficients. These approaches perform well if the phase error is low frequency. The non-parametric approaches make the assumption that all of the range data collected from a single pulse is corrupted by the same phase error, [2], [6], [7]. If this holds, nonparametric techniques are able to estimate both low and high frequency phase errors. Finally, the metric based methods generally assume that the phase error is low frequency and that it comes from a sensor velocity error. However, these methods can also be combined with non-parametric methods to produce good results, [8], [9], [10]. They utilize a cost function, such as image intensity or neg-entropy, to correct the azimuth matched filter that is applied to azimuth compress the data.What most of these methods have in common is that they must step back in the processing chain (usually azimuth decompression) to where the individual phase errors from each pulse exist to apply the phase...