An unsupervised classification method for polarimetric SAR images with a projection approach

Two unsupervised training algorithms are presented based on a projection method for polarimetric SAR images. For this, polarimetric features, including the entropy H, a-angle, anisotropy A, the largest eigenvalue of the coherence matrix lambda(1) and coherence gamma, are computed. Based on these features three noncorrelated and independent variates, which are the first three principle components are extracted using the principle component analysis. Then, a three-dimensional Cartesian coordinate is constructed using those three resultant variates. In order to reasonably define the initial cluster centers for the unsupervised classification, a projection method is proposed. As a result, a sophisticated two-dimensional feature space is obtained. Two initial cluster center determination algorithms, the histogram and histogram-quadrant methods, based on the above two-dimensional feature space are presented. Unsupervised classifications (with a minimum distance decision rule) are conducted based on the defined initial cluster centers. The classification results show that these approaches have high performance in convergence, speed, data volume and accuracy of training and classifying.