SMOOTHING PARAMETER ESTIMATION FOR MARKOV RANDOM FIELDCLASSIFICATION OF NON-GAUSSIAN DISTRIBUTION IMAGE

In the context of remote sensing image classification, Markov random fields (MRFs) have been used to combine both spectral and contextual information. The MRFs use a smoothing parameter to balance the contribution of the spectral versus spatial energies, which is often defined empirically. This paper proposes a framework to estimate the smoothing parameter using the probability estimates from support vector machines and the spatial class co-occurrence distribution. Furthermore, we construct a spatially weighted parameter to preserve the edges by using seven different edge detectors. The performance of the proposed methods is evaluated on two hyperspectral datasets recorded by the AVIRIS and ROSIS and a simulated ALOS PALSAR image. The experimental results demonstrated that the estimated smoothing parameter is optimal and produces a classified map with high accuracy. Moreover, we found that the Canny-based edge probability map preserved the contours better than others.