[GRASS-user] GRASS 7 subpixel shift of imported Landsat data

[Tomáš Brunclík]

Hello all,
working with GRASS GIS 7.0.1svn (r65198) @Linux, I noticed, that when 
importing two Landsat images from different dates, I often get them 
spatially shifted to each other when comparing them in detail - meaning 
when you zoom in to about 1:5000 scale and switch the viewed bands, the 
features in the map jump by a fraction of pixel size. For Landsat 8 
images LC81910252013208 (reflectance product) and LC81910252013224 (DN 
product) the shift is about 15m (half the pixel size) in both horizontal 
and vertical direction. The same two products imported in GRASS 6.4.4 
have no visible shift at all and the corresponding pixels seem to be 
perfectly aligned.

The procedure to reproduce this:

1)Download and uncompress this archive (133 MB):
https://dl.dropboxusercontent.com/u/24062833/pixel_shift_208_224.zip
/
The archive contains band 5 of the two above mentioned Landsat scenes, 
where the one with DOY 208 (the last number of the scene code, 
day-of-year) is from surface reflectance (SR) product, the other with 
DOY 224 is digital number (DN) product (I work on a script to 
atmospherically correct DN product based on reference SR image)//.//
/

2) Start GRASS 7, create new mapset in location with UTM 33N 
(EPSG:32633) spatial reference system.

3) Import each image, open them in the map, set colors, zoom in and 
compare a distinct feature placement in detail:

#import
r.in.gdal input="/path/to/data/LC81910252013208LGN00_sr_band5.tif" 
output=208_grass7
r.in.gdal input="/path/to/data/LC81910252013224LGN00_B5.TIF" 
output=224_grass7

#set the null pixels of the DN band (helps the bands to get similar 
color palette in the next step)
r.null map=224_grass7 setnull=0

#set colors:
r.colors map=208_grass7
r.colors map=224_grass7


4) Zoom in to a distinct feature in about 1:5000 scale (I have used a 
pond at coordinates 542365,554675), switch visibility of both images on, 
then repeatedly switch on/off the upper image in the stack. You should 
see the shift quite clearly. Zooming in even more, it is possible to 
measure the shift.

Repeat the steps 3,4 in GRASS 6 (6.4.4/Linux in my case) importing as 
208_band5_grass6, 224_band5_grass6. The result is (for me) that files 
208_grass7, 208_grass6, 224_grass6 are perfectly aligned, while 
224_grass7 is off about 15m east and 15m south from the rest (seems it 
is exactly half pixel size)...

I tried if region settings prior the import affects this, it seems not. 
Also the problem is always exactly the same when trying several times, 
creating new mapset or not.
I also tried if the images are shifted when the tiffs are opened 
directly in QGIS - they are not.
I also tried to download JPEG quicklook products of the above mentioned 
scenes (created both from DN product, JPEG color format) and the results 
are not shifted.


As I use the images in change detection, the shift causes that 
especially linear no-change features (like highways, runways) are not 
detected as no-change.

Now, any ideas why is that, and why it is so in GRASS7 and not in GRASS6?

Tomas B.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osgeo.org/pipermail/grass-user/attachments/20150527/1f3db800/attachment.html>