Statistical Classification For Polarimetric SAR Images Based On Eigenvectors

Polarimetric Synthetic Aperture Radar(POLSAR) is a multi-parameter and multi-channel Radar Imaging System, which captures polarimetric information of target by measuring the full polarimetric scattering echo in every resolution unit on the ground. Compared to the single polarization SAR, which can only detect scattering feature of target under specific combination of Polarimetric Transceiver of electromagnetic wave, the full polarimetric Synthetic Aperture Radar can provide more information for target interpretation. Image Classification of POLSAR is a key problem as well as a difficulty in the data application of POLSAR. This paper studies the statistical classification based on eigenvectors. The eigenvectors of the polarimetric coherent matrix contain the main polarimetric information, while the eigenvector corresponding to the largest eigenvalue, which is called largest eigenvector for short, is considered to be predominant. Therefore, this article studies the statistical distribution of the eigenvectors and provides a relevant classification method. The main works as follows:1)Taken as the classification characters, the largest eigenvector is transformed to a new space, in which the largest eigenvector is performed in a simple form. In this space, a three-dimensional Gaussian distribution model is given to fit the distribution of the largest eigenvector, and then, a Bayesian classifier is chosen to do the initial classification. To improve classification accuracy, a local Wishart classifier with neighborhood information is needed to carry out the second classification. The results prove that the model of the eigenvectors is effective.2)Classification in areas, where different terrain has a similar scattering machine, is not very good using the largest eigenvector as the characters. To solve this problem, a scattering angle is extracted from all the eigenvectors and eigenvalues. The largest eigenvector and the scattering angle are used together in a statistical frame to guide the terrain classification. Besides, an azimuth is extracted from the whole eigenvectors and eigenvalues as the character to make the classification of the heterogeneous better. It turns out that both methods improve the classification.