Automatic unsupervised change detection using multi-temporal polarimetric SAR data

This paper addresses the change detection capabilities of fully polarimetric synthetic aperture radar (SAR) for the L-band frequency in comparison with single- and dual-polarization and fully polarimetric SAR data. All polarization combinations are investigated quantitatively for unsupervised change detection under different topographic characteristics. In particular, highly urbanized areas, vegetated areas, and mixed topographic areas are examined. This allows optimal selection of polarization combinations that provide the highest change detection accuracy. The unsupervised change detection method applied in this study is based on a closed-loop process. Firstly, adaptive iterative filtering is used to determine the optimal filter size such that the speckle noise is sufficiently reduced. Secondly, the log-ratio image is generated from filtered SAR images and is modeled according to a Gaussian distribution. Thirdly, the Kittler-Illingworth minimum error thresholding (KI) algorithm is applied under generalized Gaussian assumptions to select double thresholding that discriminates the positively and negatively changed areas from the unchanged areas. Experimental results reveal that the combined cross-polarized (HV+VH) power data are preferable if fully polarimetric data are unavailable. The selection of filter size affects the change detection accuracy, and is dependent on the topographic characteristics.