Extracting Tidal Stream Networks from High-resolution Remote Sensor Imagery: a Preliminary Study

In this paper we report the results of a pilot project we conduct to develop a semi-automated approach for tidal stream network extraction from remote sensor data. Our test site is part of the North Inlet-Winyah Bay National Estuarine Research Reserve in Georgetown, South Carolina, U.S.A. Largely lower than 1m above sea level, the study site consists of salt marshes behind barrier islands with numerous tidal creeks. The primary data are a high-resolution multispectral image acquired with an Airborne Data Acquisition and Registration (ADAR) 5500 digital camera system in October 2000. Our working procedural route comprises four major components: pattern recognition, spatial modeling, computer vision, and interactive editing. We used pattern recognition to separate stream channels from land, and further processing focused on the stream channels only. We computed a Euclidian distance raster and a Euclidian direction raster by using the channel banks as the source. We used a non-directional filter to extract the strong edges from the Euclidian direction raster. Then, we extracted the initial stream networks (channel centerlines) by selecting the pixels with strong edge values and non-zero distance values that fall within the tidal channels. Finally, we converted the initial raster stream networks into vector polylines for interactive editing with the ADAR image as the reference in a geographic information system (GIS) environment. The result is quite encouraging as virtually all major stream systems and many small streams were successfully extracted. The limitations are mainly with some small streams whose centerlines were not preserved or were broken.