Evaluation of Polarimetric SAR Decomposition methods for tropical forest analysis

Abstract

Information about tropical forest can be obtained by remote sensing, using either optical instruments or an active radar like synthetic aperture radar(SAR). Polarimetric decomposition theorems break polarimetric SAR measurements into components that describes the scattering behavior of the target. This thesis deals with evaluating the suitability of the various decomposition theorems to describe and classify areas of tropical forest. High resolution images provided by an optical spaceborne instrument is used as ground truth information. These images are used to determine classes for segmenting the polarimetric image, and picking training and testing data for the classification procedure. The thesis focuses on multivariate Gaussian classifiers engaging the parameters associated with the components of the polarimetric decomposition theorems. There are two main goals of the project, the first is to provide a ranking to which polarimetric decomposition theorem is the best fit to describe this kind of landscape and the second is to find an optimal subset of the polarimetric features. It is shown that using compositions of polarimetric features from the decomposition theorems increases accuracy significantly compared to a classification based on intensities. Methods are first used on one test site to find an optimal composition of features, then the same features are used on another test site to prove that the composition will be effective on another site as well.