[GRASSLIST:7870] Re: visualize DEM error

[David Finlayson]

Remember that there are (at least) two sources of error:

1) Error inherent in your input data (survey error, projection
transformations, datum adjustments, etc.)

2) Error introduced by your interpolation (overshoots, undershoots,
wrong surface model, etc.)

v.surf.rst provides several tools for visually checking errors
introduced by your interpolation, including building a map of
deviations from known values that result by adjusting the
interpolation parameters (tension, smoothing, etc.) and
cross-validation which checks to see how well the model does in areas
where you do not have data (see the devi and cvdev options). Also, if
you do an r.info on your raster you will find a note in the history
section that reports the rms deviation from the input data. That is
the average magnitude of difference between your surface interpolation
and the input data.

Deviations from the known data can be removed by changing the
interpolation parameters. You can force the surface to pass through
the points/lines by setting smoothing to zero and raising the tension.
The cost is a less smooth surface that may be unrealistic. Checking
maps deviations can give you an idea of where the interpolation
parameters are breaking down. For example, the parameters work fine in
flat areas but don't work in hilly sections of your data (or
vis-a-vis). However, this map doesn't tell you above errors in you
data where you do not have values (like between contours). To estimate
these errors you use cross validation.

Cross validation works by randomly dividing you input data into two
sets a large training set, and a small test set. You run the
interpolation on the training set and compare the surface values
against the test set. The errors between the training surface and the
test set give you an indication of how well you are modeling areas
where you have no data. You should cross-validate on many different
randomly selected test sets to give you some confidence in the error
estimates. If your data are too few to remove any points or lines and
still have a reliable surface, cross-validation errors will be large.
This is an indication that your surface is likely to be unreliable in
areas where you have little data. This situation is common when
working with manually produced elevation data. It is simply too time
consuming to collect exhaustive data sets. Alternatively, you may find
that removing points has introduces little error. This means you
surface is reliable. You may have the option to thin the data and save
on computing time. (this situation is more common now that we have
high-resolution mapping tools like LIDAR and multi-beam SONAR).

Another option is to use Kriging for interpolation which uses
statistical techniques to estimate the error associated with surface
predictions. I think that kriging can be done using the R/GRASS
interface. This technique requires more expertiese than the canned
tools might suggest, however, and is less strait-forward than
cross-validation.

Be forwarned, however, that none of these techniques measure error
associated with #1 above. Garbage in equals garbage out. You can't
make a surface more accurate than the input data. Survey errors are
very common with contour data, unfortunately, since most of the
existing data was produced photogrammatically.

I live in Washington State (USA). Our mountains are covered by dense
conifer forests making the use of photogrammatic techniques difficult,
particularly picking out streams and other subtle features. When we
first introduced LIDAR mapping techniques (which can map through the
forest canopy) we expected to find elevation errors since the old
photo techniques couldn't see under the trees, what we didn't expect
to find was the artistic license that the old mappers had taken with
the topography. When ever they marked a stream into the map, they put
a little v-indent (all streams carve little notches right?). Mountain
tops and ridges always came to a point, etc. Lidar maps of the same
areas were almost foreign looking because nature wasn't so organized!

You would not be able to pick out those kinds of errors using
cross-validation or Kriging. The only way is an independent data set
like the LIDAR data we used.

David

On 8/10/05, orkun <temiz at deprem.gov.tr> wrote:
> Dylan Beaudette wrote:
> 
> > Couple ways that are outlined in this book [1] suggest making
> > histograms of elevation and aspect data. if you see spikes at regular
> > intervals in either graph then you have some artifacts - possibly due
> > to the use of contours as your input data. You can do this with the
> > GRASS-R interface in a couple of minutes.
> >
> > Another quick check, does your DEM support realistic flow patterns,
> > when modeled with something like r.terraflow? i.e. do numerous sinks
> > and dams cause water to move with no outlet, when expert knowledge of
> > the area suggests otherwise? Comparison with an alternate source of
> > stream data may also shed some light on the quality of the DEM.
> >
> > How did you create the DEM, and from what type of source ?
> >
> > [1]    J. P. Wilson and J. C. Gallant, editors. Terrain Analysis. John
> > Wiley and Sons Inc., New York, 2000.
> >
> >
> > --
> > Dylan Beaudette
> > Soils and Biogeochemistry Graduate Group
> > University of California at Davis
> > 530.754.7341
> >
> >
> > On Aug 10, 2005, at 12:46 AM, orkun wrote:
> >
> >> hello
> >>
> >> how  can  I visualize DEM error ?
> >> how can I interpret whether DEM which I have created is correct ?
> >>
> >>
> >> could you give me hints ?
> >>
> >> regards
> >>
> >> Ahmet Temiz
> >>
> >> ______________________________________
> >> XamimeLT - installed on mailserver for domain @deprem.gov.tr
> >> Queries to: postmaster at deprem.gov.tr
> >> ______________________________________
> >> The views and opinions expressed in this e-mail message are the
> >> sender's own
> >> and do not necessarily represent the views and the opinions of
> >> Earthquake Research Dept.
> >> of General Directorate of Disaster Affairs.
> >>
> >> Bu e-postadaki fikir ve gorusler gonderenin sahsina ait olup, yasal
> >> olarak T.C.
> >> B.I.B. Afet Isleri Gn.Mud. Deprem Arastirma Dairesi'ni baglayici
> >> nitelikte degildir.
> >>
> >>
> >>
> >
> >
> >
> thank you Dylan
> 
> I used v.surf.rst to create DEM.
> I have a topographic contour map as source.
> 
> regards
> 
> 
> ______________________________________
> XamimeLT - installed on mailserver for domain @deprem.gov.tr
> Queries to: postmaster at deprem.gov.tr
> ______________________________________
> The views and opinions expressed in this e-mail message are the sender's own
> and do not necessarily represent the views and the opinions of Earthquake Research Dept.
> of General Directorate of Disaster Affairs.
> 
> Bu e-postadaki fikir ve gorusler gonderenin sahsina ait olup, yasal olarak T.C.
> B.I.B. Afet Isleri Gn.Mud. Deprem Arastirma Dairesi'ni baglayici nitelikte degildir.
> 
> 


-- 
David Finlayson
Marine Geology & Geophysics
School of Oceanography
Box 357940
University of Washington
Seattle, WA  98195-7940
USA

Office: Marine Sciences Building, Room 112
Phone: (206) 616-9407
Web: http://students.washington.edu/dfinlays