Hi,
I am running GRASS 6.5 svn (Snow Leopard). I have several ASTER GDEMs (Latlong, 30m res.), which I reprojected into UTM using both ‘nearest’ and ‘cubic’ methods (‘r.proj’). Only in some imagery that are reprojected in cubic, I ended up having couple of cells (literally, 1-2 cells out of 8 million in average) with minus (-) values. For example, in a DEM where the elevation values are between 800-2600 meters, I have cell values between -150 and -85 meters. This does not seem to be an issue in reprojected imagery with the nearest method. I am not sure how these minus values are introduced at the first place.
Since I interpolate GDEMs from 30m to 10m using ‘r.resamp.rst’ (keeping tension at 40.), I need to fix this problem. I thought of making negative value cells “null” but this may create more problems during interpolation (actually, I tried and r.resamp.rst could not complete interpolation). If I take absolute values of cells to make negative cells positive, then I introduce cells with unrealistic elevation data. It seems like after interpolation –without any correction of negative values–, the number of negative value cells do not increase, they are still less than 1% of total cells.
I will appreciate any suggestions why cubic method produces negative values in some imagery and how to fix this before or after interpolation.
- I did not try changing default r.resamp.rst values, such as tension, because most of the imagery I interpolated have very accurate results; the elevation difference between interpolated (10m res) and original (30m res) is around 1 meter.
Thank you for your comments.
–
BÜLENT ARIKAN
School of Human Evolution and Social Change
Arizona State University
Tempe - AZ
85287-2402
Bulent Arikan wrote:
I am running GRASS 6.5 svn (Snow Leopard). I have several ASTER GDEMs
(Latlong, 30m res.), which I reprojected into UTM using both 'nearest' and
'cubic' methods ('r.proj'). Only in some imagery that are reprojected in
cubic, I ended up having couple of cells (literally, 1-2 cells out of 8
million in average) with minus (-) values. For example, in a DEM where the
elevation values are between 800-2600 meters, I have cell values between
-150 and -85 meters. This does not seem to be an issue in reprojected
imagery with the nearest method. I am not sure how these minus values are
introduced at the first place.
Cubic interpolation can introduce overshoot, as can other forms of
spline interpolation. Linear and nearest-neighbor interpolation don't
have this issue.
With r.resamp.rst, the problem can be alleviated to a degree by using
higher values for the tension= parameter.
Also, if your data is noisy, this will tend to exaggerate the
gradients, making overshoot more likely. Filtering the data first will
reduce the errors.
--
Glynn Clements <glynn@gclements.plus.com>
Thanks Glynn.
That makes sense now. I did not know about the inherent problem of overshooting in cubic interpolation. ASTER DEMs I have seem to be pretty consistent in terms of data ranges and representation of values (based on r.report), so I am not sure about the noise effect in my case. However, does this mean that I am sort of out of “corrective” options if I want to use cubic reprojection of imagery from Latlong to UTM? I mean I either use nearest method for error-free reprojection or I have to be careful about the results of cubic method. Finally, when you say “use higher values for tension parameter”, I guess you mean values lower than default 40. such as 30. or 20.
Bulent
On Fri, May 7, 2010 at 5:08 AM, Glynn Clements <glynn@gclements.plus.com> wrote:
Bulent Arikan wrote:
I am running GRASS 6.5 svn (Snow Leopard). I have several ASTER GDEMs
(Latlong, 30m res.), which I reprojected into UTM using both ‘nearest’ and
‘cubic’ methods (‘r.proj’). Only in some imagery that are reprojected in
cubic, I ended up having couple of cells (literally, 1-2 cells out of 8
million in average) with minus (-) values. For example, in a DEM where the
elevation values are between 800-2600 meters, I have cell values between
-150 and -85 meters. This does not seem to be an issue in reprojected
imagery with the nearest method. I am not sure how these minus values are
introduced at the first place.
Cubic interpolation can introduce overshoot, as can other forms of
spline interpolation. Linear and nearest-neighbor interpolation don’t
have this issue.
With r.resamp.rst, the problem can be alleviated to a degree by using
higher values for the tension= parameter.
Also, if your data is noisy, this will tend to exaggerate the
gradients, making overshoot more likely. Filtering the data first will
reduce the errors.
–
Glynn Clements <glynn@gclements.plus.com>
–
BÜLENT ARIKAN
School of Human Evolution and Social Change
Arizona State University
Tempe - AZ
85287-2402
Bulent Arikan wrote:
> Cubic interpolation can introduce overshoot, as can other forms of
> spline interpolation. Linear and nearest-neighbor interpolation don't
> have this issue.
>
> With r.resamp.rst, the problem can be alleviated to a degree by using
> higher values for the tension= parameter.
>
> Also, if your data is noisy, this will tend to exaggerate the
> gradients, making overshoot more likely. Filtering the data first will
> reduce the errors.
That makes sense now. I did not know about the inherent problem of
overshooting in cubic interpolation. ASTER DEMs I have seem to be pretty
consistent in terms of data ranges and representation of values (based on
r.report), so I am not sure about the noise effect in my case. However,
does this mean that I am sort of out of "corrective" options if I want to
use cubic reprojection of imagery from Latlong to UTM? I mean I either use
nearest method for error-free reprojection or I have to be careful about the
results of cubic method.
All methods of interpolation introduce artifacts; you just have to
choose which sort of artifacts you prefer. Nearest-neighbour
interpolation produces extreme gradients at sample points and zero
gradient elswhere. Linear interpolation produces extreme curvature at
sample points and zero curvature elsewhere.
Too low (or even negative) values from undershoot aren't any more
inaccurate than too high values from overshoot; they're just more
likely to be noticed (there are several places on earth where negative
altitudes would be correct).
Finally, when you say "use higher values for
tension parameter", I guess you mean values lower than default 40. such as
30. or 20.
Higher tension values cause less overshoot at the expense of greater
variations in curvature.
--
Glynn Clements <glynn@gclements.plus.com>