Automated Extraction of Drainage Channels and Roads through Deep Learning
The National Map provides geospatial data that support various uses such as resource management, disaster response, and science investigations. To properly support these needs, data themes of the National Map must be regularly updated and spatially integrated as the features on the ground change because of environmental or man-made events. The elevation theme of the National Map is managed through the 3D Elevation Program (3DEP), which is currently (2019) coordinating collection of high resolution (HR) elevation data for the United States over an eight-year period (Sugarbaker et al. 2014). Through this program, lidar point-cloud data are being collected for the conterminous United States, Hawaii, and U.S. territories, with coarser resolution interferometric synthetic aperture radar (ifsar) data being collected for the remote areas of Alaska. HR digital elevation models (DEMs) can be generated at 1 and 3 meter resolution from the lidar point-cloud data and are also furnished by 3DEP.This research develops automated methods to update the hydrography and transportation themes of the National Map in a manner that integrates with the HR elevation and image layers. Surface water drainage networks can be extracted from a HR DEM using flow-direction and flow-accumulation modelling, but results of these methods vary depending on environmental conditions and the existence of anthropogenic features that may affect the accuracy of the elevation model, such as vegetative cover, roads, bridges, and other urban structures. Hydrologic conditioning or enforcement of a HR DEM overcome some of these issues and improve flow modelling for drainage network extraction through techniques such as filtering (Passalaqua et al. 2010), filling sinks (Tarboton 1997), cutting channels through embankments at culvert and bridge locations (Poppenga et al. 2012), or burning-in existing streams (Maidment 1996). However, drainage network extraction results can vary substantially with these techniques and the methods generally require some manual intervention and/or tuning of parameters (Poppenga et al. 2013). Consequently, additional work is needed to streamline and further automate such methods for the various landscape conditions within the United States.