Polarimetric SAR segmentation based on region merging and spectral clustering

Polarimetric synthetic aperture radar (SAR) is often applied to segmentation and classification of terrain. Practically, vastness of scenes monitored requires faster and more efficient algorithms, while complexity of terrain necessitates the capabilities of coping with heterogeneous regions. To solve the above-mentioned problems, an approach of polarimetric SAR segmentation is proposed based on region merging and spectral clustering. Firstly, an over-segmentation step is enforced based on region merging, which creates many small regions with similar statistical properties. Then, spectral clustering is applied on these over-segmented regions. Using several datasets, this method has low computational complexity compared with conventional spectral clustering of pixels and is more adapted for heterogeneous terrain and large-scene compared with traditional region merging method.