Removing low noise from Semi-Global Matching (SGM) Points

At PhoWo and INTERGEO 2015 rapidlasso was spending quality time with VisionMap who make the A3 Edge camera that provides fine resolution images from high altitudes and can quickly cover large areas. Under the hood of their LightSpeed software is the SURE dense-matching algorithm from nframes that turns images into photogrammetric point clouds. We were asked whether LAStools is able to create bare-earth DTM rasters from such points. If you have read our first, second, or third blog post on the topic you know that our asnwer was a resounding “YES!”. But we ran into an issue due to the large amount of low noise. Maybe the narrow angle between images at a high flying altitude affects the semi-global matching (SGM) algorithm. Either way, in the following we show how we use lascanopy and lasheight to mark low points as noise in a preprocessing step.

We obtained a USB stick containing a 2.42 GB file called “valparaiso_DSM_SURE_100.las” containing about 100 million points spaced 10 cm apart generated by SURE and stored with an (unnecessary high) resolution of millimeters (aka “resolution fluff”) as the third digit of all coordinates was always zero:

las2txt -i F:\valparaiso_DSM_SURE_100.las -stdout | more
255991.440 6339659.230 89.270
255991.540 6339659.240 89.270
255991.640 6339659.240 88.660
255991.740 6339659.230 88.730

We first compressed the bulky 2.42 GB LAS file into a compact 0.23 GB LAZ to our local hard drive – a file that is 10 times smaller and that will be 10 times faster to copy:

laszip -i F:\valparaiso_DSM_SURE_100.las ^
       -rescale 0.01 0.01 0.01 ^
       -o valparaiso_DSM_SURE_100.laz ^

Then we tiled the 100 million points into 250 meter by 250 meter tiles with 25 meter buffer using lastile. We use the new option ‘-flag_as_withheld’ to mark all buffer points with the withheld flag so they can be easily removed on-the-fly via the ‘-drop_withheld’ command-line filter (also see the README file).

lastile -i valparaiso_DSM_SURE_100.laz ^
        -tile_size 250 -buffer 25 ^
        -flag_as_withheld ^
        -odir valparaiso_tiles_raw -o val.laz
250 meter by 250 meter tiling with 25 meter buffer

250 meter by 250 meter tiles with 25 meter buffer

Before processing millions to billions of points we experiment with different options to find what works best on a smaller area, namely the tile “val_256750_6338500.laz” pointed to above. Using the workflow from this blog posts did not give perfect results due to the large amount of low noise. Although many low points were marked as noise (violett) by lasnoise, too many ended up classified as ground (brown) by lasground as seen here:
excessive low noise affects ground classification

excessive low noise affects ground classification

We use lascanopy – a tool very popular with forestry folks – to compute four BIL rasters where each 5m by 5m grid cell contains the 5th, 10th, 15th, and 20th percentile of the elevation values from all points falling into a cell (also see the README file):
lascanopy -i val_256750_6338500.laz ^
          -height_cutoff -1000 -step 5 ^
          -p 5 10 15 20 ^
The four resulting rasters can be visually inspected and compared with lasview:
lasview -i val_256750_6338500_*.bil -files_are_flightlines
comparing 5th and 10th elevation percentiles

comparing the 5th and the 10th elevation percentiles

By pressing the hot keys <0>, <1>, <2> and <3> to switch between the different percentiles and <t> to triangulate them into a surface, we can see that for this example the 10th percentile works well while the 5th percentile is still affected by the low noise. Hence we use the 10th percentile elevation surface and classify all points below it as noise with lasheight (also see the README file).
lasheight -i val_256750_6338500.laz ^
          -ground_points val_256750_6338500_p10.bil ^
          -classify_below -0.5 7 ^
          -odix _denoised -olaz
We visually confirm that all low points where classified as noise (violett). Note the obvious “edge artifact” along the front boundary of the tile. This is why we always recommend to use a buffer in tile-based processing.
points below 10th percentile surface marked as noise

points below 10th percentile surface marked as noise

At the end of the blog post we give the entire command sequence that first computes a 10th percentile raster with 5m resolution for the entire file with lascanopy and then marks all points of each tile below the10th percentile surface as noise with lasheight. When we classify all points into ground and non-ground points with lasground we ignore all points classified as noise. Here are the results:

DTM extracted from SGM points despite low noise

DTM extracted from dense-matching points despite low noise

corresponding DSM with all buildings and vegetaion included

corresponding DSM with all buildings and vegetaion included

Above you see the generated DTM and the corresponding DSM. So yes, LAStools can create DTMs from points that are result of dense-matching photogrammetry … even when there is a lot of low noise. There are many other ways to mix and match the modules of LAStools for more refined workflows. Sometimes declaring all points below the 10th percentile surface as noise may be too agressive. In a future blog post we will look how to combine lascanopy and lasnoise for a more adaptive approach.

:: compute 10th percentile for entire area
lascanopy -i valparaiso_DSM_SURE_100.laz ^
          -height_cutoff -1000 -step 5 ^
          -p 10 ^

:: tile input into 250 meter tiles with buffer
lastile -i valparaiso_DSM_SURE_100.laz ^
        -tile_size 250 -buffer 25 ^
        -flag_as_withheld ^
        -odir valparaiso_tiles_raw -o val.laz

:: mark points below as noise
lasheight -i valparaiso_tiles_raw/*.laz ^
          -ground_points valparaiso_DSM_SURE_100_p10.bil ^
          -classify_below -0.5 7 ^
          -odir valparaiso_tiles_denoised -olaz ^
          -cores 4

:: ground classify while ignoring noise points
 lasground -i valparaiso_tiles_denoised\*.laz ^
          -ignore_class 7 ^
          -town -bulge 0.5 ^
          -odir valparaiso_tiles_ground -olaz ^
          -cores 4 

:: create 50 cm DTM rasters in BIL format
las2dem -i valparaiso_tiles_ground\*.laz ^
        -keep_class 2 ^
        -step 0.5 -kill 200 -use_tile_bb ^
        -odir valparaiso_tiles_dtm -obil ^
        -cores 4 

:: average 50 cm DTM values into single 1m DTM 
lasgrid -i valparaiso_tiles_dtm\*.bil -merged ^
        -step 1.0 -average ^
        -o valparaiso_dtm.bil

:: create hillshade adding in UTM 19 southern
blast2dem -i valparaiso_dtm.bil ^
          -hillshade -utm 19M ^
          -o valparaiso_dtm_hill.png

:: create DSM hillshade with same three steps
las2dem -i valparaiso_tiles_raw\*.laz ^
        -step 0.5 -kill 200 -use_tile_bb ^
        -odir valparaiso_tiles_dsm -obil ^
        -cores 4
lasgrid -i valparaiso_tiles_dsm\*.bil -merged ^
        -step 1.0 -average ^
        -o valparaiso_dsm.bil
blast2dem -i valparaiso_dsm.bil ^
          -hillshade -utm 19M ^
          -o valparaiso_dsm_hill.png

3 thoughts on “Removing low noise from Semi-Global Matching (SGM) Points

  1. Pingback: Creating a Better DTM from Photogrammetic Points by Avoiding Shadows | rapidlasso GmbH

  2. Pingback: Integrating External Ground Points in Forests to Improve DTM from Dense-Matching Photogrammetry | rapidlasso GmbH

  3. Pingback: Removing Excessive Low Noise from Dense-Matching Point Clouds | rapidlasso GmbH

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s