Correction Methods for Non-Stationary Noise Floor in Sentinel-1 Images Using Convex Optimization

Abstract

Synthetic aperture radar (SAR) is a method of creating images of the surface of the Earth by emitting and receiving radar waves. Sentinel-1 is a SAR platform made by the European Space Agency (ESA) that provides a source of SAR images open to the public through the operation of two satellites. Due to the non-uniform radiation pattern projected from the satellite's antenna, there are significant non-stationary noise floor intensity patterns that distract from the desired measurements, which are particularly significant in certain types of image modes, namely Extra Wide and Interferometric Wide modes. While ESA provides a default noise floor estimate with each Sentinel-1 product, with the intention that it be subtracted from the original image so the result is homogeneous, there is clear evidence that it is miscalibrated. This Masters thesis presents two novel methods for estimating the noise floor patterns in the images that are demonstrated to be improvements over the default noise floor. The first method presents a way to dynamically construct and apply linear rescaling to the default noise floor estimate over different sections of the images, called subswaths, by use of least squares optimization. While the method is successful in improving image quality, it is not totally effective because the default noise floor is mis-fit in a non-linear manner. The second method constructs a new noise floor as a power function of the radiation pattern power by using linear programming and least squares optimization. This successfully compensates for the non-linear mis-fit, resulting in an overall increase in image quality, albeit with greater parametric complexity. These methods greatly improve the intrinsic value of Sentinel-1 images in scenarios where the noise floor dominates, such as in cross-polarized images and images where the physical materials result in lower backscatter intensity.