Spatial Content Understanding of Very High Resolution Synthetic Aperture Radar Images

Kurzfassung

Availability of large amounts of very high resolution (metric-resolution) remote sensing images from the last generation synthetic aperture radar (SAR) satellites is attracting new studies. In order to search and retrieve relevant images from large-scale databases, new techniques for automatically analyzing, interpreting and indexing SAR images are required. The methods developed for this purposes in the past were based on the understanding speckle characteristics in SAR images. The focus has been generally on the model-based textural parameter estimation in the amplitudeenvelope of SAR images, such as parametric Gibbs-based methods in the Bayesian framework. Such methods were largely successful under the assumption of stationarity of the signal in an analyzing window of convenient size on images with resolution of the order of tens of meters. The challenge we encounter in metric-resolution SAR images is the presence of a very high order of details encapsulating a non-stationarity, where model-based parameter estimation becomes inaccurate. This constraint encourages us to focus on nonparametric strategies while employing phase information to transform SAR images in a suitable space. Demonstrating the advantages and relevance of the phase information embedded in complex-valued SAR images over the use of the mere amplitudeenvelope for such strategies is an underlying contribution of this thesis. The importance of phase information is advocated with a proposed method of multiple sublook decomposition (MSLD). This method generates hyper-images from the spectral analysis of complex-valued SAR images enabling the visual exploration of targets. Subsequently, a chirplet-derived transform- the fractional Fourier transform (FrFT) has been found to be a true SAR relevant multi-scale approach, where scaling is carried out in the phase. A proposed non-parametric feature descriptor based on the use of second-kind statistical measures (logarithmic-cumulants) estimated over the amplitude-envelope of the FrFT coefficients exhibits enhanced feature space separability for improved indexing. An experimental benchmarking database is generated on single look complex (SLC) spotlight mode TerraSAR-X images for the validation of the proposed FrFT-based nonparametric technique in comparison to the existing methods. A robust methodological classification framework has been proposed for the evaluation and comparison of the studied algorithms.