Quantitative Comparison of Unsupervised Change Detection Capability in Multiple Polarimetric SAR Data

摘要

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, a highly urbanized area, a vegetated area, and a mixed topographic area were examined. This allows optimal selection of polarization combinations that provide the highest change detection accuracy. The unsupervised change detection method applied in this study was based on a closed-loop process. Firstly, adaptive iterative filtering was used to determine an optimal filter size such that the speckle noise was sufficiently reduced. Secondly, the log-ratio image was generated from the filtered SAR images and was modeled according to a Gaussian distribution. Thirdly, the modified Kittler-Illingworth minimum error thresholding (KI) algorithm was applied under generalized Gaussian (GG) assumption to select double threshold that discriminates the positively and negatively changed classes from the unchanged class. Experimental results reveal that the most suitable data used for the change detection was the combined cross-polarized (HV +VH) power image, because it can achieve high correct change detection rate for any topography. The selection of filter size affects the change detection accuracy, and was dependent on the topographic characteristics. In addition, the use of the combined polarized power data, which were generated after filtering the single-polarization data at each filter size, was found out to increase the change detection accuracy.