RADARSAT-2 polarimetric SAR data for urban land cover mapping using spatial- temporal SEM algorithm and mixture models

This paper presents a multi-temporal Stochastic Expectation-Maximization (SEM) algorithm with adaptive Markov Random Field (MRF) for analysis of polarimetric SAR (PolSAR) data for urban land cover mapping. The fitness of alternative distributions of multi-look PolSAR data based on Wishart, G⁰_p and K_p assumptions are compared by using the SEM algorithm. The proposed pixel-based SEM algorithm explores the spatial-temporal contextual information to improve the classification accuracy while simultaneously overcomes the pepper-salt effect of speckle. Owing to the adaptive MRF analysis, the iterative process of the supervised SEM algorithm becomes stable. Further, detailed shape features could be preserved comparing with the traditional MRF methods. Four-date RADARSAT-2 polarimetric SAR data over the Greater Toronto Area are used for the experiment. The results show that this algorithm could generate reasonable classification accuracy for detailed urban land cover mapping. And the fitness of G⁰_p and K_p distributions are proven to be better for urban land cover mapping than that of Wishart distributions.