A Novel Contextual Classification Algorithm for Multitemporal Polarimetric SAR Data

This letter presents a pixel-based contextual classification algorithm by integrating a multiscale modified Pappas adaptive clustering (mMPAC) and an adaptive Markov random field (AMRF) into the stochastic expectation-maximization process for urban land cover mapping using multitemporal polarimetric synthetic aperture radar (PolSAR) data. This algorithm can effectively explore spatiotemporal contextual information to improve classification accuracy. Using the mMPAC, the problem caused by the class feature variation could be mitigated. Using the AMRF, shape details could be preserved from overaveraging that often occurs in many nonadaptive contextual approaches. Six-date RADARSAT-2 PolSAR data over the Greater Toronto Area were used for evaluation. The results show that this algorithm outperformed the support vector machine in producing homogeneous and detailed land cover classification in a complex urban environment with high accuracy.