Unsupervised Change Detection from Multichannel SAR Data by Markov Random Fields

Abstract

In the contexts of environmental monitoring and disaster management, techniques allowing one to detect the changes that occurred in a given area between successive acquisition dates may serve as efficient data-analysis tools. Good discrimination between changed and unchanged areas can often be achieved by analyzing optical multispectral data in the related multidimensional feature space [20][43]. However, such data are affected by atmospheric and sun-illumination conditions. Synthetic aperture radar (SAR) [15][27][34] is insensitive to such issues, hence, multitemporal SAR imagery is expected to play an important role, for instance, in ecological applications [18] or in disaster assessment and prevention [11][42]. However, SAR usually allows only one amplitude/intensity observation, thus possibly resulting in poor discrimination. Multichannel (i.e., multipolarization and/or multifrequency) SAR represents an option with improved potential: as compared with single-channel SAR, it is expected to provide an increased discrimination capability, while maintaining its insensitivity to atmospheric and illumination issues. This potential is also reinforced by the availability of multichannel SAR data with short revisit times which are granted by future missions (e.g., COSMO/SkyMed and TerraSAR).