Unsupervised learning rules for POLSAR images analysis

Abstract

It has been shown (see Chitroub, S. et al., Signal Processing, vol.82, no.1, p.69-92, 2002) that the model for POLSAR (polarimetric synthetic aperture radar) images is a mixture model that results from the product of two distributions, one characterizes the target response and the other characterizes the speckle phenomenon. For scene interpretation purpose, it is desirable to separate between the target response and the speckle information. We propose here to use some unsupervised learning rules for POLSAR images analysis via a PCA-ICA neural network model. Based on its rigorous statistical formulation (see Chitroub et al., Intelligent Data Analysis International Journal, vol.6, no.2, 2002), a neuronal PCA approach for the simultaneous diagonalization of the signal and noise covariance matrices is proposed. The goal is to provide PC images that are uncorrelated and have an improved SNR. Speckle is a non-Gaussian multiplicative noise, and the higher order statistics contain additional information about it. ICA is used to separate the speckle from the PC images and providing new IC images that have an improved contrast. The method has been applied on real POLSAR images. The extracted features are quite effective for scene interpretation.