Improvement of Scattering Model Decomposition by ESPRIT-based Pol-InSAR

Observation of the global environment by microwave remote sensing technology has been attracting attention recently. Image classification is one of the important applications in POLSAR (Polarimetric SAR) image analysis. Various techniques are proposed. The scattering model decomposition [1] is one of the powerful techniques among them. This technique decomposes a covariance matrix derived by the POLSAR image data into three fundamental covariance matrices corresponding to single, doublebounce, and volume scattering component. However, when we apply the scattering model decomposition [1], we encounter problem that power of several components in some area becomes negative. This is caused by the assumptions used in the decomposition technique [1]; 1) assumption of the volume scattering matrices, and 2) one of the elements in the single or double-bounce scattering component is assumed to be known. Yamaguchi et. al. have proposed a modified technique which is referred as a conventional technique in the followings refered [2]. In this technique, the covariance matrix of the volume scattering is extended to three types. We can estimate a suitable matrix for decomposition among them. However, various types of volume scattering would exist in actual dataset. In addition, the assumption 2 is still required for this decomposition. The problem of the negative power components is improved by this technique. However there still remain several negative power areas. For the POLSAR image analysis, we only have three observables (HH, HV and VV data) so that further modification would be difficult. In this paper, we propose a modified model decomposition algorithm for Pol-InSAR (Polarimetric and Interferometric SAR) images for forest area. For the Pol-InSAR image analysis, additional images in different orbit are available. As in [5], the volume scattering component in each image area can be directly estimated for the dataset. Main cause of negative components would be selection of the inadequate covariance matrix. Therefore, this cause can be removed by the proposed technique with PolInSAR analysis. In addition, more realistic assumptions for the single and double-bounce component decomposition are considered here to obtain further accuracy improvements. Experimental results for forest analysis by the proposed and conventional technique [2] is presented to show availability of the proposed technique.