Statistics on High Resolution urban polarimetric images: Application to segmentation and classification

Inhalt:
In the field of polarimetric, interferometric, or multivariate SAR image segmentation or classification, statistical models are used to design equality test or similarity measure. In this work we address the impact of the choice of various non gaussian models on the final performance in classification or segmentation of urban polarimetric SAR images. As more parameters are used in the description of the distribution, more estimation errors are introduced. It yields that the most complex distribution performance can often be outmatched by simpler models, even in very high resolution or urban settings. As the clutter distribution evolves from gaussian noise to an impulsive non gaussian distribution, there is a breaking point at which using a non gaussian model really becomes beneficial performance wise. In this paper we present preliminary results and methods to improve the choice of an appropriate signal model for a given polarimetric SAR image.