Abstract:

In bankline extraction from remote sensing images, the results are usually rough and segmented. A new bankline extraction method based on the principal curves was proposed. In the learning process, the polygonal line (PL) algorithm and the error back propagation (BP) algorithm were combined. Firstly, the principal curve of the river centerline was learned. Then, a segmentation method was proposed to divide the riparian points into two sets which belong to the left and right bank respectively, and the smooth parameter expressions of the principal curves of the two banklines were given. Finally, the vec-tor description of the river centerline and banklines in remote sensing images were realized. The principal curve descriptions made the extracted banklines smooth, and the separate learning of the two banklines ensured the integrity of the extracted banklines for even narrow river channels. Comparison with the existing methods through experiments on real remote sensing images shows that the proposed method can achieve better smoothness and can be used to solve the problem of discontinuity in narrower channel of a river. The resulting vector descriptions of banklines are more convenient for storage and reconstruc-tion and can be used as shape features for the detection and identification of river area in images.

Key words: remote sensing image, bankline extraction, principal curves, PL algorithm, BP algorithm