The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
321
grid node can be re-validated as well. The full process may be
re-iterated as needed.
After the last iteration, each image has its own radiometric
polynomial model (eq. 1) and cloud masks can be extracted
from invalid points.
3. IMPLEMENTATION AND TESTS
3.1 Implementation
3.2.2 Influence of parameters: Radiometry is extracted
every 1000 m from sub-sampled images at 100m. To speed up
the process, images are saved in a pyramidal format so that
radiometry is interpolated directly in the right pyramid level. A
grid node is ignored if it covers global mask and invalidated if
its calculated radiometry is superior to initial threshold (180).
Here, we get 412143 valid radiometric values to set up the
global equation system. Degrees of polynomials P and Q are set
to 1, so radiometric model for a given image I is as follows (eq.
6):
This algorithm has been implemented in GeoView Software, a
photogrammetry and remote sensing suite developed in-house
at IGN-France. We focused on SPOT5-HRS images but his
module works also with any type of ortho-images or raw
images with location modelling. It also works with
multispectral images. In this case, each channel is
radiometrically adjusted independently.
3.2 Test case over Algeria
3.2.1 Dataset overview: 9 SPOT5 HRS stereo-pairs
covering a part of Algerian territory are used to test previously
described process of radiometric block adjustment. It is a
medium size dataset that can be considered as quite easy and
favourable because:
• There are large overlapping areas between images
(20-25%),
• Interval between first image taken on 09/08/2002 and
last image taken on 26/10/2002 is very short, less than
3 months,
• Very few clouds are present in images.
Radiometric adjustment is done on ortho-images calculated
from back images of HRS stereo-pairs. These ortho-images are
very large, up to 57474 columns by 115981 rows, their pixel
size is 5m and the whole dataset covers 6° of longitude by 8° of
latitude (530 km * 970 km). A global mask over Mediterranean
Sea is added.
Figure 1. SPOT5 HRS images over Algeria - Dataset overview
FinalRad(oolro\ty= [l -col+bj ■ row+C/]• InitialRadfcolran}
-\\dj • col+ej ■ row+fj ]
There are 6 unknown values by image, 54 for the whole system
that is therefore very redundant. Role of every parameter has
been studied but let’s focus on punctual constraints influence.
With Gpi=aQi=a the value used to weight punctual constraint
equations (eq. 3), Table 1 and Figures 2, 3 show huge influence
of these punctual constraints on final radiometry. With o=10,
RMS of radiometric differences on overlapping area drop from
26.5 numerical counts to 4.4, final mosaic is seamless and
image dynamic is kept. Whereas, with o=0.1 and o=l, the
system is too rigid, final radiometry close to initial one and
seams remain visible. With o=100, final image dynamic is very
poor. Standard deviation of radiometric values stored in the grid
drops to 3.3, too much information has been lost.
% valid
nodes
Grid average
radiometry
Grid std
dev
Residual
RMS
Initial
99.1
100.2
25.9
26.5
o=0.1
98.5
100.2
17.2
13.5
0=1
98.1
100.2
14.7
7.8
o=10
93.8
100.2
11.1
4.4
o=l 00
93.8
100.2
3.3
0.5
Table 1. Influence of punctual constraints
Figure 2. Influence of punctual constraints on final radiometry