Geomorphons - processing British Isles SRTM DEM with GRASS7 r.geomorphon extension

Some time ago I mentioned Geomorphons, a machine vision based approach to topographic analysis promoted by Stepinski and Jasiewicz.

If you don't know what a geomorphon is, here's a link to the Space Informatics Lab at University of Cinncinati, see also the papers: Stepinski+Jasiewicz 2011, Jasiewicz + Stepinski 2013 

At the time I first came across this I had some trouble getting GRASS7 to install and get the extension working in Linux so I tried running GRASS7 in windows inside a VM but it didn't really work well.


Now however I have a working GRASS7 installation on Linux and here's a few demonstration maps, which show the broad categories of landscape position. Geomorphons use a line of sight approach best explained by the links and papers above.

I have used a search distance of L = 1000m in these maps, using Shuttle Radar Topography Mission data for the British Isles. I've set a threshold for flatness as 0.2 degrees, which is probably too small because in flat areas I'm not confident of the reality of these interpretations.

Geomorphons have a different kind of scale-dependence than the usual differential geometry based approaches used, though it is not independent of scale, it is more so than a simple slope analysis.

Co-ordinates are OSGB all-numeric throughout.

West Cornwall

SE Cornwall

N Cornwall

 

Aberystwyth area in mid-Wales

Mid-wales the same as above but with flatness threshold of 1 degree

Changing the threshold for flatness to 1.0 degrees, flat areas such as Borth Bog are shown as such on the map.

A closeup with 5m data:

Nevertheless I keep search distance L=1000m.

Using QGIS profiler plugin to compare 2m resolution LiDAR vs SRTM

I have an idea to produce an elevation aware cycling route planner, and the first step in that is to have some elevation data in the form of a Digital Terrain Model.

The most commonly used freely available digital elevation data is that derived from the Shuttle Radar Topography Mission. At the latitude of Cornwall (about 50 degrees N) it has a resolution of about 73m x 73m per cell.

There is however a possibility of getting LiDAR data (free for non-commercial use) from the UK Environment Agency, which has various resolutions, and I obtained some 2m resolution data for an area around Truro.

For a cycling route planner, it would be nice to be able to keep track of the smaller bumps that would be smoothed out by the SRTM data. However it might not be practical to use the high resolution data because once one considers larger areas the volume of data becomes very large indeed.

There is also the Terrain 50m raster data available through the Ordnance Survey website.

I wished to see what difference it would make on a simple test circular route in Truro. I chose one which starts off climbing a hill, then goes along relatively flat terrain and then back down the hill again by a different route.

The QGIS screenshot below shows using the "Profile Tool" QGIS plugin, where the darker shade of green is the lower elevation terrain. You may need to click on the image to view it in a larger format.

The actual profile itself showing the elevation as a function of distance along the path is as follows:

SRTM = red, OS Terrain50 = black, 2m LiDAR = green

 I have had an issue with the QGIS Profile tool, in that I can't currently get it to work from a saved track, rather than 'live' by point and click each time, therefore the 2nd image above is a slightly different path.

Unfortunately the LiDAR has some wildly oscillating values, particularly around 400m from start where the road is in a cutting near the top of Chapel Hill. I have therefore done a moving average smoothing to make them more comparable.

Although the OS Terrain 50m is a higher resolution dataset, it appears to have more artifacts than the SRTM which is ~73m at this latitude

Measuring total ascent, counting the total height gain considering only positive values:

Not a big difference in total ascent between SRTM and Terrain 50. I also made an estimate by inspection of a 1:25k OS map of 55m ascent.

The total ascents were 52m, 55m, 60m, and 75m respectively for the smoothed LiDAR, SRTM, Terrain 50, and unsmoothed LiDAR respectively.

In the early part of the track, the SRTM (red line) appears to be seeing the treetops or roofs of buildings either side of Chapel Hill, but levels out at a lower peak than the full resolution LiDAR (yellow line). It appears than the LiDAR is overestimating due to proximity of artificial structures, or possibly even vehicles in a traffic jam on the A390. The smoothed LiDAR follows the SRTM fairly closely.

An elevation aware cycling route planner?

For some time now, I've had an idea to produce an elevation aware cycling route planner. You see, Google Maps can give you a route, but it doesn't take account of hills when deciding it.

Basically the idea I have is based on attaching some kind of cost distance multiplier to segments of route.

One idea I have had is to use the RSGISLib tools to segment a layerstacked digital elevation model, consisting on the elevation, slope, aspect (degrees from N) and then intersect this with a map of the road network (it is possible to download versions of this derived ultimately from OpenStreetMap and the OS OpenData has data available as well) to make road segments that have a consistent slope.

Then the average slope in the direction of the road could be calculated, and from this a cost-distance multiplier.

This is a segmented DEM in mid-Cornwall from the SRTM and with a pixel size of 73m, with a minimum object size of 9 pixels.

I haven't yet worked out how to complete this, one important thing is that the relevant slope is that in the direction of the road rather than the absolute value, and of course it is different if traversed in the reverse direction. So perhaps it would be necessary to convert every road into two one-way roads, so that the cost-distance could be calculated separately for each direction

Dissertation done, and the likelihood of Martian glaciers in Cornwall, Wales and south-west England

So, the dissertation is done now, you can get a copy here, or a special tablet-optimised version here. The latter is using the US Supreme Court's paper size and Century Schoolbook font, as was advised to me by http://en.wikibooks.org/wiki/LaTeX/Page_Layout.

In the dissertation itself, I explain that the Martian terrain is segmented using RSGISLib and a classifier function assigned to the segments using the Souness et al. 2012 glacier-like forms as a guide.

So what if we apply it to a terrain that is not on Mars? The Shuttle Radar Topographic Mission data is available at a similar resolution to that from HRSC for Mars.

I have done the segmentation for areas of Wales, Cornwall and the southwest of England using topography only rather than integrating the red image field, since it wouldn't really be comparable with all the vegetation etc. on the Earth.

The same procedure of highlighting segments with log(K) > 12, 13, 14, 15 with a semi-transparent overlay is used. The background images are from Landsat 8 using bands 6, 5, 2 for RGB. All images from 25th July 2014.

Ordnance Survey GB numerical coordinates are used.

Not much in West Cornwall, except on the north-facing slopes west of St. Ives. There are a few more segments on the North Cornwall coast and in some areas of southeast Cornwall.

The north coast of West Penwith has a high likelihood of Martian glaciers, see also the map in the following post which shows a relatively shallow area of sea that would have been dry land in the Early Holocene, perhaps sediment deposited by the Martian glaciers at some point.

Some martian glaciers expected on the northern fringe of Dartmoor near Okehampton, but perhaps surprisingly also in south Devon.

Exmoor seems a favoured location for martian glaciers.

The Welsh valleys and the Brecon Beacons are also highlighted for a high likelihood of martian glaciers.

Since Wales generally has a higher likelihood of martian glaciers, I have used a slightly different scaling with only starting to highlight at log(K) > 13:

A close up of the Valleys and Brecon Beacons:

Mid-Wales:

North Wales:

Snowdonia:

Aberystwyth area:

The southern part of Cwm Rheidol appears particularly favoured for martian glaciers.

Segmenting with LandSerf: Wales DEM

Using some more SRTM tiles, I will try segmenting the DEM of Wales, using the elevation, slope, aspect, and several curvature layers generated in LandSerf:

 Here is an example segmentation showing the mid-Wales area with RSGISlib with a minimum object size of 80 pixels (0.2sqkm):

Segmenting Topography with LandSerf and RSGISLib

As a training for segmenting Martian topography, I have done a little work on Earth.

Using the software LandSerf  by Prof. Jo Wood (the website www.landserf.org appears to be down at the moment), I created a number of derived topograhic layers from Shuttle Radar Topography mission data, i.e. slope, aspect, and plan, longitudinal,  cross-sectional, profile and mean curvature. I then layerstacked these files (taking the aspect in degrees from north-facing).


Using RSGISlib to segment these (using min 128px sized objects where a pixel is ~ 72m) I get results like this:

The curvature layers seem to enable it to segment along ridge and valley lines and it can sometimes be seen how the topography is orientated spatially.

Self-paced Introduction to QGIS and more on DEM segmentation

I have seen recently there is a self-paced course on QGIS from the  FOSS4Geo Academy . It is hosted on the open source Canvas Learning Network.

There are two modules so far, an introductory one, and a cartography one, with more coming soon.

Following on from my last post, here is the topography of the British Isles segmented using RSGISLib using a minimum object size of 65536 pixels.

This probably isn't a useful approach, since it took the computer about 9 hours to do this. I expect it is better to use smaller objects and aggregate them at the classification stage.

Here the mean segments are colourised using yellow for slope, and blue for elevation with a gaussian stretch as before.

 And here are the segments with a random colourisation.

Using a higher resolution (5m) DEM for mid-Wales, around the Aberystwyth, Dyfi estuary and Pumlumon area, aggregating to at least 1024 pixels (that is equivalent to a 160x160m square):

Object-based segmentation of topography with RSGISLib

The RSGISLib (Remote Sensing and GIS) Python library written by Pete Bunting and Daniel Clewley provides a number of features, among them segmentation of images to objects.

This can not only be applied to what one conventionally thinks of as images, but also topographic layers. I thought I'd practice this on Earth first before trying to do this for Mars as I will be for my dissertation.

Using Space Shuttle Radar data (freely available - I got it via the program Viking) I obtained the elevation of the British Isles (excluding Shetland) and using GDAL via QGIS and the command line, created slope and aspect layers. I then transformed the aspect to remove the discontinuity at 360-0 deg, using degrees from N, so that 180 is south facing, east and west are both 90 etc.

Using RSGISLib it is possible to segment to objects. Here I show some results of doing so, using a layerstacked image with elevation, slope and angle from N. I have set the minimum connected object size to 1000 pixels. With the topographic data gridded at 73m, this means objects about 5.5 sq. km or larger.


Visualisation is a gaussian stretch in TuiView, setting red to aspect, green to slope, and blue to elevation, showing mean values for each segment.

Now I show a few of the actual segments without the values, with colours assigned in RSGISLib for visualisation. It is possible to populate a Raster Attribute Table to add data to the segments for a land cover classification.

From the Isles of Scilly to the Isle of Wight:

Wales, and central England.

 Zooming in to mid Wales:

By changing the object size, it is possible to do this at different scales. This is part of the segmentation for a minimum object size of 16384 pixels (that is about 89 sq. km) equivalent to a square 9.5km a side:

The thing that makes geospatial data interesting is how what you see in it changes depending on what scale you look at it and how you visualise it.