Maximum likelihood approach to the detection of changes between multitemporal SAR images

The authors introduce maximum likelihood techniques for optimised discrimination between agricultural and wooded regions, based on a multitemporal sequence of ERS images. The inherent resolution of the system is inadequate to make such a classification on an individual image. However, the different temporal change patterns of the two classes can be exploited. One approach uses joint annealed segmentation of the image sequence, providing optimised exploitation of the speckle model in determining the common set of region boundaries in the underlying radar cross-section. This is followed by maximum likelihood change detection using a normalised log temporal texture measure. This is shown to be superior to constructing the normalised log measure directly, including speckle fluctuations, followed by a single annealed segmentation process. Finally, it is demonstrated how simple filtering of this normalised log measure can provide reasonable classification with greatly reduced computation load.