We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

eess.IV

Change to browse by:

References & Citations

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Electrical Engineering and Systems Science > Image and Video Processing

Title: Unsupervised Classification for Polarimetric SAR Data Using Variational Bayesian Wishart Mixture Model with Inverse Gamma-Gamma Prior

Abstract: Although various clustering methods have been successfully applied to polarimetric synthetic aperture radar (PolSAR) image clustering tasks, most of the available approaches fail to realize automatic determination of cluster number, nor have they derived an exact distribution for the number of looks. To overcome these limitations and achieve robust unsupervised classification of PolSAR images, this paper proposes the variational Bayesian Wishart mixture model (VBWMM), where variational Bayesian expectation maximization (VBEM) technique is applied to estimate the variational posterior distribution of model parameters iteratively. Besides, covariance matrix similarity and geometric similarity are combined to incorporate spatial information of PolSAR images. Furthermore, we derive a new distribution named inverse gamma-gamma (IGG) prior that originates from the log-likelihood function of proposed model to enable efficient handling of number of looks. As a result, we obtain a closed-form variational lower bound, which can be used to evaluate the convergence of proposed model. We validate the superiority of proposed method in clustering performance on four real-measured datasets and demonstrate significant improvements towards conventional methods. As a by-product, the experiments show that our proposed IGG prior is effective in estimating the number of looks.
Subjects: Image and Video Processing (eess.IV)
Cite as: arXiv:2104.01656 [eess.IV]
  (or arXiv:2104.01656v1 [eess.IV] for this version)

Submission history

From: Shijie Ren [view email]
[v1] Sun, 4 Apr 2021 18:10:30 GMT (4403kb,D)

Link back to: arXiv, form interface, contact.