Urban Land Cover Classification of Polarimetry Images Using Adaptive Contextual SEM Approach
Mahdianpari, Masoud¹; Motagh, Mahdi^2
1University of Tehran, IRAN, ISLAMIC REPUBLIC OF; ²GFZ, GERMANY

Land cover classification with Polarimetric Synthetic Aperture Radar (PolSAR) is a challenging task in urban areas Remote Sensing. In this paper we propose a parametric pixel-based methodology for efficiently classifying PolSAR data to drive land cover products in urban areas. One of the problems associated with PolSAR classification is the existence of inherently multiple speckle noise which influences the results of terrain classification. We address this issue using a combined pixel-wise and contextual analysis for the representation of restored SAR data. A probability density function (PDF), which is fitted well with the data based on goodness-of-fit, is first obtained for pixel-wise analysis. Then contextual smoothing is achieved with the Markov Random Field (MRF). Consequently speckle-free images are achieved by weighted summation of PDF and MRF outcomes.
Following speckle reduction, an unsupervised contextual analysis is used to improve the pixel-based image classification by exploring spatial information available in the neighboring graphical elements. For this reason a MRF-based approach is used which works on the Bayesian frame with Markovian dependency assumption. The traditional MRF analysis using stable neighborhood structures and fixed contextual parameters usually lead to over-averaged results, where detailed urban features disappear. To solve this problem we propose a novel approach using an adaptive MRF with adaptive parameters and neighborhood. We employ the stochastic expectation maximization (SEM) algorithm to jointly perform clustering of the data and parameter estimation of the statistical distribution. This distribution conditioned to each image clusters and the MRF model. Our proposed classification methodology is applied on medium resolution L-band ALOS data from Tehran, Iran and the results are compared with other commonly used approaches.