An analysis on the contribution of the phase component for scene class recognition in TerraSAR-X scenes

Traditional approaches for SAR image classification and target recognition in high resolution acquisitions discard phase information and are solely based on the detected data. When one makes use of several SLC image acquisitions, the phase variation is the main source of information. We propose a method for feature definition and extraction which combines single SAR acquisitions with interferometric information. Not only that phase becomes less random, but it describes the changes that occur on the received scene reflectivity. In the hypothesis that two successive acquisitions are taken on a stationary scene (tandem acquisitions), combined information from these data can be employed to describe the scene content. Thus, we propose the usage of spectral based descriptors that make use of the full complex signal from two or more SLC acquisitions to construct robust image descriptors. The experiments show that the interferometric information can be very valuable for the recognition of coherent targets, increasing the recognition rate. Our approach relies on the estimation of the SAR image spectra, from which features are derived in two stages: the estimation of the model order and the extraction of the most relevant descriptors. The set of descriptors is extracted from the complex spectrum of the SAR image, using a spectral decomposition method. We consider the problem of estimating the parameters of complexvalued two-dimensional sinusoidal signals observed in noise. Our approach considers not just isolated objects but also takes into account the complexity of high resolution urban scenes, where there is a large variety of structures, which can be best understood in their context. This is why we propose to extract the feature vectors from image patches which cover a relevant area on the ground (200 x 200 m), compared to the object size. Experiments are performed on data acquired in the context of scientific proposals LAN 1613, MTH 1628, and OTHER0502