Urban Mapping Using Multitemporal Very High Resolution Sar Data by a Knowledge-based Sem Algorithm

The objective of this research is to assess the multitemporal very high resolution single polarization SAR data for urban land cover/land-use mapping using a novel knowledge-based SEM algorithm. Three-date RADARSAT-2 ultra-fine beam SAR data were collected over the rural-urban fringe of Greater Toronto Area. A modified Stochastic Expectation-Maximization (SEM) algorithm which employs an adaptive Markov Random Field (MRF) and the Finite Mixture Model (FMM) was proposed for the supervised classification. Several SAR intensity distribution models such as Gamma, K, G0 and Fisher were compared using the algorithm. A set of rules according to the diversity of the land cover texture patterns was further applied in the decision fusion to improve the urban land cover classification. Preliminary results show that the proposed algorithm which explores the spatio-temporal information with the knowledge about the ultra-high urban SAR textures could produce reasonable classification results. Homogeneous urban land cover maps could be obtained while the detailed shape features could be preserved. Although the overall classification accuracy of the single polarization data set is not as high as desired, more details could be identified in the very high resolution SAR data. Using unique very high resolution SAR textures, rules were designed to effectively improve the classifications of several land cover classes thus improve the overall classification accuracy.