Context-Aware Filtering of PolSAR Data Based on Quaternion Wavelet Shrinkage

Kurzfassung

Polarimetric synthetic aperture radar (PolSAR) is widely used for land cover classification. One of its major problems is the effect of speckle. Although numerous methods have been developed to suppress speckle noise, it is difficult to classify complex land cover types composed of multiple different scattering mechanisms by pixel-wise classification with existing filtering methods. In this paper, we propose context-aware PolSAR filters based on quaternion wavelet shrinkage to solve this problem. We first represent logarithmic-transformed Stokes vectors (log Stokes vectors) in a quaternion form. Next, we apply the quaternion wavelet transform to the quaternion images of log Stokes vectors. Then, we perform filtering in the quaternion wavelet domain by shrinking small wavelet coefficients. Log Stokes vectors in areas with highly random polarimetric information, i.e., areas where the polarization state varies from pixel to pixel, are strongly filtered, while those in homogeneous areas are weakly filtered or unfiltered. Therefore, after the inverse transform, filtered log Stokes vectors incorporate spatial-context information as the degree of polarization (DoP). This results in accurate classification of complex land cover types without using contextual classifiers or a large number of texture features. Experimental results show that the proposed filters are more suitable than conventional filters for the preprocessing of pixel-wise land cover classification.