MultiSeg: A hierarchical segmentation algorithm for radar and optical data
ISBN 978-85-88783-11-9
Authors
1Costa, G.; 2Dutra, L.; 3Uba, D.; 4Soares, M.; 5Feitosa, R.; 6Rosa, R.
1PUC-RIO Email: gilson@ele.puc-rio.br
2INPE Email: dutra@dpi.inpe.br
3INPE Email: douglas@dpi.inpe.br
4INPE Email: marinalva@dpi.inpe.br
5PUC-RIO Email: raul@ele.puc-rio
6BRADAR Email: rafael.rosa@bradar.com.br
Abstract
This paper describes the underlying algorithm and key implementation aspects of a hierarchical segmentation program called MultiSeg Built on a previous development called SegSar, MultiSeg consists on a specialized segmentation technique devised for SAR (Synthetic Aperture Radar) and optical imagery. Initially, images are compressed at different rates creating an image pyramid, then a region growing procedure is used in combination with a split and merge technique at the different compression levels. In sequence, the program processes the image pyramid from the coarser to the finer compression levels, applying a border refinement heuristic each time it changes from one level to the next. MultiSeg was created in the C++ language, with the support from the open-source library TerraLib. The devised software architecture permits easy extension of its capabilities. Additionally, preliminary tests have shown that MultiSeg is capable of processing large volumes of data efficiently. This paper also presents an objective evaluation of segmentation results obtained with MultiSeg, produced through comparing the automatic segmentation with manually delineated reference segments.
Keywords
remote sensing; radar; segmentation