Improving Multispectral Image Classification by Using Maximum Pseudo-Likelihood Estimation and Higher-Order Markov Random Fields

In this paper we address the multispectral image contextual classification problem following a Maximum a Posteriori (MAP) approach. The classification model is based on a Bayesian paradigm, with the definition of a Gaussian Markov Random Field model (GMRF) for the observed data and a Potts model for the a priori knowledge. The MAP estimator is approximated by the Game Strategy Approach (GSA) algorithm, a non-cooperative game theory based method. Maximum Pseudo-Likelihood is adopted for MRF model parameter (regularization parameter) estimation on higher-order neighborhood systems. To evaluate the proposed method, experiments using Nuclear Magnetic Resonance (NMR) images were proposed. Quantitative results obtained by using Cohen’s Kappa coefficient for performance evaluation show significant improvement in contextual classification, indicating the effectiveness of our MAPMRF approach. Keywords-Markov Random Fields;Contextual classification; Maximum pseudo-likelihood; Game Strategy Approach.