The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Publications Copernicus
Articles | Volume XL-1/W3
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XL-1/W3, 257–261, 2013
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XL-1/W3, 257–261, 2013

  24 Sep 2013

24 Sep 2013


M. Mahdian1, S. Homayouni2, M. A. Fazel1, and F. Mohammadimanesh1 M. Mahdian et al.
  • 1Dept. of Geomatics Engineering, College of Engineering, University of Tehran, Iran
  • 2Dept. of Geography, University of Ottawa, Canada

Keywords: Unsupervised Classification, Expectation Maximization (EM), Polarimetric Synthetic Aperture Radar, Mellin transform, Markov Random Field (MRF)

Abstract. The discrimination capability of Polarimetric Synthetic Aperture Radar (PolSAR) data makes them a unique source of information with a significant contribution in tackling problems concerning environmental applications. One of the most important applications of these data is land cover classification of the earth surface. These data type, make more detailed classification of phenomena by using the physical parameters and scattering mechanisms. In this paper, we have proposed a contextual unsupervised classification approach for full PolSAR data, which allows the use of multiple sources of statistical evidence. Expectation-Maximization (EM) classification algorithm is basically performed to estimate land cover classes. The EM algorithm is an iterative algorithm that formalizes the problem of parameters estimation of a mixture distribution. To represent the statistical properties and integrate contextual information of the associated image data in the analysis process we used Markov random field (MRF) modelling technique. This model is developed by formulating the maximum posteriori decision rule as the minimization of suitable energy functions. For select optimum distribution which adapts the data more efficiently we used Mellin transform which is a natural analytical tool to study the distribution of products and quotients of independent random variables. Our proposed classification method is applied to a full polarimetric L-band dataset acquired from an agricultural region in Winnipeg, Canada. We evaluate the classification performance based on kappa and overall accuracies of the proposed approach and compared with other well-known classic methods.