Riemannian Clustering of PolSAR Data using the Polar Decomposition
Résumé
In this manuscript we propose an algorithm 1 for unsupervised classification of PolSAR data, on the 2 manifold of Hermitian positive definite matrices obtained 3 from the polar decomposition of the scattering matrix. The 4 method uses a geodetic metric for evaluating similarity of 5 Hermitian matrices and performs unsupervised classifica-6 tion for both coherent and incoherent targets. Monostatic, 7 full-polarimetric, real and simulated datasets are used for 8 testing the proposed method. With Gaussian clutter, the 9 technique is able to retrieve classification maps similar to 10 those obtained using the standard Wishart algorithm. A 11 refinement of classification results is shown for a simulated 12 dataset with 4 classess. While the Wishart classifier attains 13 an average class accuracy of almost 97%, the proposed 14 method reaches almost 99%. For real PolSAR data, the 15 final classification better preserves the texture information 16 of the original image. As a result, an improved separation 17 is shown between nearby areas of lower intensity, as for 18 example vegetation fields. 19
Origine | Fichiers produits par l'(les) auteur(s) |
---|---|
Licence |
Copyright (Tous droits réservés)
|