Linear Feature Extraction from High-Resolution SAR Images

The raising availability of Synthetic Aperture Radar (SAR) images makes it possible to acquire and add important information to the one obtained via the classical optical domain. Nevertheless, SAR images are not so easy to interpret because they are subjected to distortions (i.e. foreshortening, layover, and radar shadows) and they show a disturbance usually referred as multiplicative speckle noise. In an automatic image interpretation system, the accurate extraction of the main features, which are present on the acquired scene, is usually the most fundamental step to deal with. In this context, linear feature extraction plays an important role for a fully automatic image understanding system. In this thesis, linear feature extraction issue is firstly tackled by detecting amplitude discontinuities (i.e. edges) of the signal. Next, the obtained result is refined by an edge linking stage so that, finally, the boundary of the object of interest is extracted throughout a higher-level model. Exploiting the speckle model in the edge detector operation, the knowledge of the probability density function (pdf) of the filtered data is used to obtain a desired probability of false alarms (PFA) at the output of the system. In this context, both parametric and non-parametric statistical edge detectors are evaluated on highresolution SAR images. Next, presupposing classical assumptions on the distribution of such images, a novel statistical model is devised to cope with the final pdf of general linear filtered data. Even though the presented model is an approximation of the real pdf that does not have closed form, it fits well the data and its mathematical treatability enable us to exploit the model in different way, even improving the performance of the subsequent linking stage. To free the processing by statistical assumptions of data, general multiscale linear filtering is investigated. Several methods of scale combinations and automatic thresholding are evaluated to obtain an improvement on the final edge detection performance. The edge linking stage presented in this thesis makes use of the sequential edge linking (SEL) algorithm already known in literature, which models the linking issue as a shortest path problem (SPP). Nevertheless, the application of the original method to the one-look SAR data shows many problems and drawbacks. For this reason, two novel SPP metrics have been proposed and new steps have been added to the original algorithm. Following the study on the edge detection, both a parametric and a non-parametric metric has been proposed to make the method generally applicable. Furthermore, exploiting the model developed for the edge detection, we proved the useful of its application in improving SEL performance. The algorithm devised to extract linear features (e.g. roads and runways), relies on the Hough Transform framework and tries reconstructing the object boundaries as composition of linear segments. In particular, in the extraction step, the road is recognized as composition of regions that are limited by parallel lines and characterized by a low radar cross section (RCS) due to a high homogeneous material. Moreover, to study the loss of information occurring in the discretization of the Hough Transform definition, new theoretical bounds on the parameter sampling have been obtained and compared with their signal processing counterpart. Finally, to improve edge detection performance and approach a higher level of processing, a novel despeckling algorithm is presented. It belongs to the class of non-linear, anisotropic diffusion filters that apply a partial differential equation (PDE) onto the image in analysis. In particular, since the proposed filter is a PDE-based filter, no noise model is presupposed so that, in principle, it can be applied to any noise type. Consequently, no statistical modeling effort is required to change sensor or data type (e.g. intensity or amplitude). Ultimately, since the application of the proposed despeckled filter enables a straightforward segmentation to be applied, even on one-look Cosmo-SkyMed images, such approach extends the extraction of a possible general shaped feature.