River network completion without height samples using geometry-based induced terrain

We present an improved method for connecting broken river segments by incorporating segment geometry into the induced terrain approach. The reconnection problem is important in aerial photography. Canopies and clouds cover parts of rivers, and a complete river network is necessary for transportation, land-use planning, and floodplain control. Our previously presented induced terrain approach guarantees that the global network topology is hydrologically consistent. However, it assumes many reliable height samples, which may be unavailable because of extensive view obstacles or gentle surfaces, to reconstruct the induced terrain surface. In this paper, we exploit river segment geometries to generate the induced terrain. We assign lowest heights to the known river locations, and higher and higher heights for locations further and further away. Since areas radiating from segment tips tend to be where a river is broken, we assign relatively small heights to these areas, so as to favor water flow toward them in the subsequent river derivation. On average, such a tip-biased model improves reconnection accuracy by five percentage points (which is 40\% of what we would have been corrected if we had had rich height samples covering 10\% of the terrain cells) over the intuitive baseline model in which the height of a location increases with its distance from the respective nearest river location, without considering the potential effect of segment tips as the tip-biased approach.