High resolution radargrammetry: development and implementation of an innovative image matching strategy

Synthetic Aperture Radar (SAR) is a microwave imaging system that, installed as payload on satellite platform, offers some important advantages considering its complete independence from logistic constraints on the ground (as for airborne data collection), illumination (daylight) and weather (clouds) conditions. One of the most important applications of SAR remote sensing is the generation of Digital Surface Models (DSMs), that is, three-dimensional models of the Earths surface; these 3D models have a large relevance in many engineering, environmental, civil protection, safety and security applications. Starting from the SAR data two different methods may be used to generate DSMs: the well-known interferometric (InSAR) and the radargrammetric one (StereoSAR), and the importance of the latter approach is rapidly growing just due to the new high resolution imagery (up to 1 m of GSD - Ground Sample Distance). Radargrammetry in principle requires just a couple of images to reconstruct the three-dimensional shape of the Earth, like in the human stereo vision or in the classical photogrammetry applied to optical imagery; on the contrary, it does not need the good coherence among two (or a series of) images, a strict requirement for the success of the interferometric approach which can be guaranteed only with an extremely short (tens of seconds) revisit time. This study is focus on the present potentialities of high resolution SAR satellite imagery for DSMs generation with a radargrammetric stereo-mapping approach. Therefore, the aim of this thesis is methodological, devoted to illustrate both the fundamental advantages of this approach and also its drawbacks. In detail, this thesis was focused on the crucial step for DSMs generation, which is the image matching. The matching process is the automatic identification of pixels representing the same object in the two (or even more) images. If corresponding pixels are recognized, than a simple geometric intersection is needed to compute the three-dimensional position of the object in space. In particular, the main goal of this study was to look up the answer to the following research questions: • How can an image matching strategy be improved and adapted to the complex high resolution SAR imagery in order to exploit their potentiality for the radargrammetric DSM generation? • What are the necessary key steps to perform an image matching procedure and to developed a complete radargrammetric processing chain? The development of a fully automatic, robust and reliable image matching method that adapts to different images and scene contents is a challenging problem. Dissimilarities between SAR imagery due to occlusion, geometric distortions, radiometric differences and speckle noise must be take into account. An matching algorithm is generally composed of two essential parts: select a primitive model to identify a correspondence between the pixels of the two or more images and define a search strategy to find the matching candidates. The developed algorithm, presently under patenting by the University of Rome "La Sapienza", is based on an hierarchical solution with a geometrical constrain and the correspondences are looked using an area based matching primitive model and analysing the signal to noise ratio (SNR). Moreover, experimental results have highlighted that an image enchantment should be consider; in this work different speckle filter methods have been investigated and embedded in the radargrammetric software. A complete radargrammetric processing chain has been developed and in order to demonstrate its mapping potentialities, several tests were carried out using high resolution SAR satellite imagery with different acquisition mode (SpotLight, StripMap) and coming from different platforms (COSMO-SkyMed, TerraSAR-X). Four test sites, Merano, Como (Northern Italy), Gloggnitz (Eastern Austria) and San Francisco (California, USA) have been selected in order to highlight for each of them peculiarity, difficultly and particularity detected during the processing: • Merano test site: analysis of foreshortening and layover effect on DSM accuracy and comparison between Stripmap and Spotlight imagery • Como test site: analysis of radargrammetric potentiality over complex urban areas and study of advantages using ascending and descending stereo-pairs • Gloggnitz test site: the developed radargrammetric processing chain has been compared to the in-house DLR ones, directly by the Astrium-Infoterra image analysts • San Francisco test site: a comparison between DSM generated with InSAR and StereoSAR techniques has been performed within a collaboration with e-Geos spa (an Italian company) and sarmap sa (a Swiss company) A homogeneous DSM assessment procedure has been considered in the different tests carried out; it is based on a comparison with a reference ground truth using the scientific software DEMANAL (developed by K. Jacobsen - Leibniz University of Hannover) and the accuracy statistics are compute at 95% probability level. Summarizing, the DSM vertical accuracy is strictly related to the terrain morphology and land cover; the observed RMSE values range from 3-4 meters over bare soil and forest to 6-7 meters in more complex urban areas. In zones affected by strong SAR geometric distortions (Layover and Foreshortening) the terrain morphology could be conveniently reconstructed using two same-side stereo pairs acquired from different look side. Finally, radargrammetric stereo-mapping approach appears a valuable tool to supply topographic information and is likely to became an effective complement/alternative to InSAR technique, since it may work using a couple of images with a good performance even over areas characterized by low values of coherence.