|
|
| GeoCommunity Mailing List |
| |
| Mailing List Archives |
| Subject: | GISList: SUM re: determining DEM accuracy or error |
| Date: |
04/03/2003 07:05:00 PM |
| From: |
northms .. telus.net |
|
|
Detailed SUM/Notes below.
thanks,
sarah
Sarah North
northwest hydraulic consultants
email. northms@telus.net
ORIGINAL QUESTION:
Can anyone give suggestions about how to report accuracy of a DEM and
measurement derived from a DEM?
I have one DEM each for the same study area in two separate years. In each
case, one-metre contours and additional planimetric data were used to build a
TIN elevation model. The TIN's were converted to grid-type DEM's with two-
metre cell size. The two surfaces were compared by subtracting the year1 DEM
from the year2 DEM.
How do I go about estimating the accuracy (or error) of each DEM, and the
accuracy of difference measurements taken from the two DEM's (i.e., year2 -
year1)? Can anyone suggest methodology or references to help with this?
SUMMARY/NOTES:
Thanks to several GISList respondents: Kevin Murphy, Holly Glaser and
Jayachandran Mani. Also thanks to my colleague, Darren Ham, who provided
several suggestions and references. I still have a lot to learn about this,
but below are my notes on the subject, so far.
Re: accuracy of the input DEM's:
- Kevin Murphy sent links for mapping standards (see below).
- Holly Glaser & Jayachandran Mani both sent suggestions (see below).
- Ideally, elevation accuracy should be reported by the data supplier. In my
case, the original contour and planimetric data, mapped using photogrammetric
techniques, were accurate to half a contour interval or better in unobscured
areas, to one contour interval or better in areas obscured by trees. TIN's
were derived from these data, then the TIN's were converted to DEM's for
volumetric measurement.
- Darren Ham reminded me to consider cell size when creating a grid from
contour or spot elevation data. In general, the larger the cell size, the
larger the potential error introduced. Use cell size similar to the average
spacing of raw data when creating the grid from spot elevation data. When
creating the grid from contour data, use a cell size comparable to the minimum
distance between contours.
Re: accuracy of measurements taken from the two DEM's (i.e., year2 - year1):
- Consider horizontal accuracy. I.e., if two elevation surfaces are being
compared and they are not geometrically registered to each other, that may
introduce errors in the vertical difference measurements. This was not an
issue for me, as the data for year2 were mapped using year1 data as a base.
- Calculate root mean square error (RMSE). E.g., When measuring stream bed
sediment volume change from year1 to year2):
elevation difference RMSE (per cell)
= sqrt [ (year1 error)^2 + (year2 error)^2 ]
= sqrt [ (2 x 0.5)^2 + (2 x 0.5)^2 ]
= 1.4
where year1 error = 2 x the accuracy, which in this case is +/- 0.5
metres
and where "^2" means "squared"
volume difference RMSE (per study area) = (study area) x (elevation
difference RMSE per cell)
- This value may be very large in relation to your measured net sediment
volume (e.g., several times the measured number), suggesting that the measured
value is not significantly different than measurement error. However, this is
a worst-case value: it is quite likely that accuracy is really much better than
this. You might want to calculate RMSE using an assumed typical input data
accuracy of +/- 0.1 m instead of 0.5 m. Even so, in my case the numbers were
very large in relation to measured net volume.
- Another method (courtesy of Darren Ham): Find areas that you think should
not have changed from year1 to year2. Calculate the net change in volume
between the two surfaces for these areas, and divide by area to get an estimate
of volume error per unit area. Multiply this by the measured values for the
entire study area. I couldn't use this approach in my project, as the year2
data were mapped with year1 data as a base, and only areas of change were
remapped.
- Another measurement to try: Overlay the point data from which the surface
was derived with the derived surface. For each point, get a difference between
the original and interpolated value. RMSE = sqrt [ sum (original -
interpolated)^2 ]
- Jack-knifing: Hold back some of the original data points. Compare these
points to the interpolated surface to estimate RMSE.
- For more on this subject (I just got these references today, haven't read
them yet), see U.S. Army Corps of Engineers' papers:
Byrnes, M.R. et al. 2002. "Quantifying Potential Measurement Errors and
Uncertainties Associated with Bathymetric Change Analysis" - available as PDF
online at: http://chl.wes.army.mil/library/publications/chetn/pdf/chetn-iv50.pdf
Johnston, S. 2002. "Uncertainty in Bathymetric Surveys"
|
|
Sponsored by:
For information
regarding
advertising rates
Click Here!
|
