Experimenting with Hydrological Analysis using TauDEM

Blue Rivers Ordered

Over the past few years, I’ve played around with developing ordered rivers networks for different projects. I am not an expert in hydrology, but I can get close for cartographic purposes. I am an expert; however, in asking for help from those who know best and I rely on a lot of very smart people to guide me on my journey.

Recently, I decided to put together a visualization of ordered rivers for New Zealand. I came across a very nice data set offered through the Ministry for the Environment via the Koordinates website and thought I’d like to put it to use.

The rivers vis project made me wonder if I could build this base dataset myself using some of the recently released elevation data sets via the LINZ Data Service. The short answer to my question is “sorta”. Doing it open source is not an issue, but building an accurate ordered river centerline network is another story. This is a task I cannot take on as a solo project right now, but I could do a little experimentation. Below, I’ll offer some of methods and things I learned along the way.

Tools and Data

The method I tested used TauDEM and a 1m DEM raster accessed from the LINZ Data Service. I down sampled the DEM to 2m and 5m resolutions and used small areas for testing. Finding and open source tool was easy. I sorted through a few available methods and finally landed on “Terrain Analysis Using Digital Elevation Models” (TauDEM). There are additional methods through GRASS and SAGA GIS. I chose TauDEM because I never used it before.

Method Tested

To my knowledge, there is no open source tool where a person can put in a DEM and get a networked rivers centerline vector out the other side. It requires a number of steps to achieve your goal.

The basic run down to process the DEM is to:

  1. Fill sinks
  2. Determine flow directions
  3. Determine watersheds
  4. Determine flow accumulation
  5. Stream classification
  6. Export to vector

TauDEM does require a few extra steps to complete the process, but these steps are explained in the documentation of the tool. It was more about keeping all my variables in the right places and using them at the right time. I recommend using the variable names TauDEM provides.

BASH script here

Click the arrow to the left to view the full BASH script below:


#Rough sketch for building river centerlines. Rasters have been clipped prior


raster_list=$( find $BASEPATH -name "*.tif" )



for i in $raster_list


	file_name=$( basename $i )

	strip_input_extension=$( echo $file_name | sed 's/.tif//' )


	gdal_translate -tr $reso $reso -of GTiff $i $reso_name.tif




	#TauDEM Commands
	mpiexec -n 8 pitremove -z $processed_input_file -fel $fel

	mpiexec -n 8 d8flowdir -fel $fel -p $p -sd8 $sd8 

	mpiexec -n 8 aread8 -p $p -ad8 $ad8 -nc

	mpiexec -n 8 dinfflowdir -fel $fel -ang $ang -slp $slp

	mpiexec -n 8 areadinf -ang $ang -sca $sca -nc

	mpiexec -n 8 slopearea -slp $slp -sca $sca -sa $sa

	mpiexec -n 8 d8flowpathextremeup -p $p -sa $sa -ssa $ssa -nc

	mpiexec -n 8 threshold -ssa $ssa -src $src

	mpiexec -n 8 streamnet -fel $fel -p $p -ad8 $ad8 -src $src -ord $ord -tree $tree -coord $coord -net $net -w $w


The script is a rough sketch, but does get results.

Challenges in the Process

One major challenge for this project was the size of the input DEM and my computers available RAM. I work primarily off a laptop. It’s a good machine but no match for a proper server set up with some spacious RAM. My laptop struggled with the large hi-resolution DEMs, so I needed to down-sample the images and choose a smaller test area to get it to work.

Clip the tiff with gdal_translate -projwin and down sample with -tr

gdal_translate -tr xres yres -projwin ulx uly lrx lry input.tif output.tif

The second challenge came up because I used a bounding box to clip my test regions. I recommend not doing this and instead clip your regions using a watershed boundary. Having square shapes for your test regions will give you very inaccurate and unhelpful results. For example, major channels in your DEM will be cut at the edges of your raster. You will not get accurate results.

Clipping a raster using a shapefile, like a watershed boundary, can be achieved using gdalwarp.

gdalwarp –cutline input.shp input.tif output.tif


I ran my process and QCed the results against Aerial Imagery and a hillshade I developed from the DEM. The first run gave me good enough results to know I have a lot of work to do, but I did manage to develop a process I was happy with. The tool did a great job, but the accuracy of the DEM was a little more challenging. It’s a start. I captured a good number of river channels despite my incorrect usage of a square DEM, learned a lot about how DEM resolution affects outputs, and gained knowledge around how to spot troublesome artifacts.

Well Defined ChannelsImg 1: River capture in well defined channel.

From this experiment, there are a few ideas I’d like to explore further:

1. Accuracy of the DEM. The particular DEM I worked with had a number of ‘dams’ in the flows. Notably, bridges, culverts, vegetation artifacts, and other general errors that caused water to flow in interesting directions. When working with a data set like this, I am curious how manage these artifacts.

Road issueImg 1: River diversion at road.

Artifact issueImg 1: River diversion at culvert or bridge.

2. How to go beyond borders. This analysis can be broken down by watershed, but it will be necessary to link the outflows of those watersheds to the next for accurate results.

Edge issueImg 1: Flow not captured at edge.

3. As DEMs are released with better resolution, there is a need for scaled up computing power. The process needs a large amount of RAM. What is the best computational set up for capturing the largest area?

4. Did I do this correctly? I perform this task about once every two years and usually weekends when the surf is flat and the garden is weeded, so I am not an expert. There is a lot more research to be done to determine if I am using the tools to the best of their abilities.

Wellington Elevations: Interpolating the Bathymetry

It is important to note something from the very beginning. The interpolated bathymetry developed in this project does not reflect the actual bathymetry of the Wellington Harbour. It is my best guess based on the tools I had and the data I worked with. Furthermore, this interpolation is NOT the official product of any institution. It is an interpolation created by me only for the purposes of visualization.


Part of the goal when visualizing the Wellington landscape was to incorporate a better idea about what may be happening below the surface of the harbor. Various bathymetric scans in the past have gathered much of the information and institutions like NIWA have done the work visualizing that data. As for myself, I did not have access to those bathymetries; however, I did have a sounding point data set to work with, so I set about interpolating those points.

The data set, in CSV format, was over a million points; too dense for a single interpolation. I worked out a basic plan for the interpolation based on splitting the points into a grid, interpolate the smaller bits, then reassemble the grid tiles into a uniform bathymetry.

Conversion from CSV to shp
Using the open option (-oo) switch, OGR will convert CSV to shp seamlessly

ogr2ogr -s_srs EPSG:4167 -t_srs EPSG:4167 -oo X_POSSIBLE_NAMES=$xname* -oo Y_POSSIBLE_NAMES=$yname*  -f "ESRI Shapefile" $outputshapepath/$basenme.shp $i

Gridding the Shapefile
With the shapefile in place, I next needed to break it into smaller pieces for interpolation. For now, I create the grid by hand in QGIS using the ‘Create Grid’ function. This is found under Vector>Reasearch Tools>Create Grid. Determining a grid size that works best for the interpolation is a bit of trial and error. You want the largest size your interpolation can manage without crashing. Using the grid tool from QGIS in very convenient, in that it creates an attribute table of the xmin, xmax, ymin, ymax corrodinates for each tile in the grid. These attributes become very helpful during the interpolation process.

Interpolating the Points
I switched things up in the interpolation methods this time and tried out SAGA GIS. I have been looking for a while now for a fast and efficient method of interpolation that I could easily build into a scripted process. SAGA seemed like a good tool for this. The only drawback, I had a very hard time finding examples online about how to use this tool. My work around to was to test the tool in QGIS first. I noticed when the command would run, QGIS saved the last command in a log file. I found that log, copied out the command line function, and began to build my SAGA command for my script from there.

Here is look at the command I used:

saga_cmd grid_spline "Multilevel B-Spline Interpolation" -TARGET_DEFINITION 0 -SHAPES "$inputpoints" -FIELD "depth" -METHOD 0 -EPSILON 0.0001 -TARGET_USER_XMIN $xmin -TARGET_USER_XMAX $xmax -TARGET_USER_YMIN $ymin -TARGET_USER_YMAX $ymax -TARGET_USER_SIZE $reso -TARGET_USER_FITS 0 -TARGET_OUT_GRID "$rasteroutput/sdat/spline_${i}"

I tested a number of methods and landed on ‘grid_spline’ as producing the best results for the project. It was useful because it did a smooth interpolation across the large ‘nodata’ spaces.

Once the initial interpolation was complete, I needed to convert the output to GeoTIFF since SAGA exports in an .sdat format. Easy enough since GDAL_TRANSLATE recognizes the .sdat format. I then did my standard prepping and formatting for visualization:

gdal_translate "$iupput_sdat/IDW_${i}.sdat" "$output_tif/IDW_${i}.tif"
gdaldem hillshade -multidirectional -compute_edges "$output_tif/IDW_${i}.tif" "$ouput_hs/IDW_${i}.tif"
gdaladdo -ro "$output_tif/IDW_${i}.tif" 2 4 8 16 32 64 128
gdaladdo -ro "$ouput_hs/IDW_${i}.tif"2 4 8 16 32 64 128

Here is look at the interpolated harbour bathymetry, hillshaded, with Wellington 1m DEM hillshade added over top

And here is a look at the same bathy hillshade with coloring

Visualizing the Bathymetry
With the bathymetry, complete it was simply a matter of building it into the existing visualization I built for the Wellington Region. Learn more about the project here. The visualization was four steps:

Aerial Imagery
Then merge the models together

Easy as, eh? Let me know what you think!

Note: All imagery was produced during my time at Land Information New Zealand. Imagery licensing can be found here:
“Source: Land Information New Zealand (LINZ) and licensed by LINZ for re-use under the Creative Commons Attribution 4.0 International licence."