A semisupervised contextual classification algorithm for multitemporal polarimetric SAR data

This paper presents a contextual classification algorithm which employs the multiscale modified Pappas adaptive clustering (MPAC) approach and the Semisupervised Expectation-Maximization (SEM) procedure for urban land cover mapping using multitemporal polarimetric SAR (PolSAR) data. The proposed pixel-based algorithm explores spatio-temporal contextual information and thus could effectively improve the classification accuracy while simultaneously avoids the pepper-salt results which often occurs on the SAR images. Moreover, owing to the multiscale analysis, MPAC could adaptively preserve the detailed features comparing with other non-adaptive contextual methods. The proposed algorithm is computationally efficient and requires less parameter to be estimated. Properties of the proposed algorithm including the MRF impact, multiscale efficiency, computational performance and the initialization influence were investigated. Six-date RADARSAT-2 polarimetric SAR data over the Greater Toronto Area were used for validation. The results show that this algorithm could generate homogenous and detailed mapping results with fair accuracy for complex urban land cover classification.