A knowledge-based, two step procedure for extracting channel networks from noisy DEM data

Abstract We present a new procedure for extracting channel networks from noisy DEM data. The procedure is a knowledge-based, two-step procedure employing both local and nonlocal information. In particular, we employ a model of an ideal drainage network as a source of constraints that must be satisfied by the output of the procedure. We embed these constraints as part of the network extraction procedure. In a first step, the procedure employs the facet model of Haralick to extract valley information from digital images. The constraints employed at this stage relate to conditions indicating reliable valley pixels. In a second step, the procedure applies knowledge of drainage networks to integrate reliable valley points discovered into a network of single-pixel width lines. This network satisfies the constraints imposed by viewing a drainage network as a binary tree in which the channel segments have a one-pixel width. The procedure performs well on DEM data in the example investigated. The overall worst-case performance of the procedure is O ( N ) log N ), but the most computationally intensive step in the procedure is parallelized easily. Hence the procedure is a good candidate for automation.