Segmentation of SAR images using multitemporal information

The maximum likelihood method of SAR segmentation has the potential to retain single pixel accuracy without requiring heuristic decisions. Normally, a probabilistic measure is used to merge individual regions without assuming any prior knowledge for the underlying cross-sections. However, for a reasonable multitemporal scene, there may be considerable information available from the varying cross-sections over time. An example is given where this information can be extracted by an initial classification. It is then shown how the segmentation scheme can be modified to incorporate this information via an estimate of the multitemporal underlying class distributions. Using single-look Radarsat data at 8 m resolution, it is demonstrated how the final segment population can be significantly reduced. From a comparison with ground survey data and a high-resolution AirSAR image, the structural quality of the segmentation is shown to be improved.