Abstract
Synthetic aperture radar (SAR) is an active radar often mounted on airborne platformslike airplanes or on satellites, where it transmits electromagnetic waves to the measured
scene and records the echoes. Due to the nature of the transmitted waves, it is capable
of acquiring data regardless of weather or lighting conditions. Therefore, many Earth
Observation applications can benefit from SAR, such as environmental monitoring, ship
detection, and so on.
However, the data acquisition process inevitably incorporates phase errors, mainly
resulting from inaccuracies in the measurement of SAR platform trajectory, and these
phase errors further lead to defocusing effects in the formed SAR images. Additionally,
speckle noise always appears in the formation of SAR images due to coherent summation of the received waves. Consequently, autofocusing and despeckling approaches are
indispensable to the improvement of SAR image quality.
To achieve SAR autofocusing, we treat it as an inverse problem and minimize a
cost function regularized by a Cauchy penalty. Since the latent SAR image and the
phase errors are both unknown, the strategy of alternating minimization is adopted.
Under this framework, we propose two novel approaches for SAR autofocusing. They
estimate the phase errors in the same way, but deal with the complex nature of the image
reconstruction step differently. The first method relies on Wirtinger calculus and a linear
approximation of non-linear equations. The second one relies on a complex version of
the well-known forward-backward splitting method. The convergence and extension for
both methods are discussed, and experimental validation is done on multiple datasets,
including simulated scenes and real SAR images. Experimental results demonstrate that
both proposed methods produce SAR images with well-focused targets and outperform
the state-of-the-art.
For SAR speckle reduction, a new despeckling algorithm inspired by the wide use of
deep learning and the trend of combining model-based optimization algorithms with datadriven approaches that use neural networks is presented. A logarithmic transformation
followed by a wavelet transform is applied to the speckle image, and the obtained noisy
coefficients are denoised by four deep unfolded networks. The structure of each network
replicates the behaviour of forward-backward splitting method that minimizes a cost
function with Cauchy regularization. Experiments on both synthetic and real images
demonstrate a solid despeckling performance and a decent generalization ability of the
proposed method, in that it can remove speckle effectively under diverse circumstances.
i
Besides, it performs better than learned iterative shrinkage thresholding algorithm
(LISTA), and is comparable to the forward-backward splitting method.
Date of Award | 1 Oct 2024 |
---|---|
Original language | English |
Awarding Institution |
|
Supervisor | Alin Achim (Supervisor) & Odysseas Pappas (Supervisor) |