On the Use of TanDEM-X Bistatic InSAR Images for Scene Recognition

Kurzfassung

Research on SAR interferometry is mainly focused on the two main application areas, namely the generation of accurate digital elevation models and change detection, and also on the crucial processing steps like interferogram filtering, phase unwrapping, co-registration etc. This study rather aims to make use of interferometric SAR (InSAR) images, once they are generated, for scene recognition purposes. In literature, limited research is available on the use of InSAR images for object recognition and scene classification. For those studies, the main trend is in the direction of using the interferometric coherence for mostly binary classification such as forest/non-forest, urban/non-urban or change/no-change classification. On the other hand, available research on multi-class classification is mostly based on the temporal variation of interferometric coherence and/or backscatter intensity. However, in our studies, we emphasize the use of whole complex-valued InSAR image for multi-class classification, i.e., recognition of more complex scenes such as forest, agricultural fields, water body, different kinds of residential and industrial areas, etc., and also their combinations. Furthermore, a new complex-valued phase-gradient InSAR (PGInSAR) image is defined whose phase represents the magnitude of the phase gradient of InSAR in range and azimuth directions. Interferometric phase being related to the terrain height, phase of PGInSAR image can be considered as a measure of the terrain slope, i.e., how fast the interferometric phase changes over the image. This review study serves as a comparative assessment of various feature descriptors such as Gabor-based, FrFT-based, BoVW-based and partial derivatives based features extracted from the complex-valued InSAR and PGInSAR images, and used for patch-based classification. For this purpose, an image patch database is generated from bistatic interferometric pairs acquired by the TanDEM-X mission over the test site Toulouse, France. In order to investigate the impact of effective baseline on the classification, 3 datasets are constructed from the bistatic acquisitions with 3 different effective baselines over the same area. Supervised KNN classification is performed on the image patches of 200 x 200 pixels from 8 different scene classes representing different natural and man-made structures on a 400m x 340m terrain (Figure 1).