Extracting an accurate river network: Stream burning re-revisited

Abstract

Extracting river networks that are both accurate and topologically connected is important for applications that involve correct routing of material, for example water and sediment, through such networks. We combined water and sediment extraction using radar and multispectral imagery from Sentinel-1 and Sentinel-2 to create both water and sediment masks over a range of study areas. These were then used to condition topographic Digital Elevation Models (DEMs) by lowering the elevation of pixels with both water and sediment present, in a process known as stream burning. We examined how stream burning could improve accuracy of extracted networks and identified the most effective method of burning for optimal results. We find deeper burning depths improved accuracy, with diminishing returns: we suggest burning 40 to 50 meters. We find sediment burning improves accuracy in humid and temperate landscapes, but arid landscapes should be burned using only water pixels. We find accuracy of extracted networks is significantly better on the COP30 global topographic dataset compared to the NASADEM dataset, mainly due to the time of collection. The AW3D30 DEM and FABDEM datasets have accuracies just below that of the COP30 DEM.