MMPümsffi
jtv ■.
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
in order to secure the performance of in-flight image quality.
This processing line will be implemented under the
characteristics of the PHR system for an experimental in-flight
commissioning checking, on a specific campaign which will
amend by command the mechanical characteristics and generate
vibrations.
First, an overview of data and background will be shown.
Second, we will explain the principles of the new local
integration method and its results. Third, the PHR system
characteristics for this processing application are explained.
And fourth we will give the performance of the algorithm on
several cases of simulation for the PHR system, which is
broadly satisfactory.
Image 2 :« maître »
Image 2 :« secondaire »
Terrain
(MNT)
As it will be shown further, the new integration method is
applicable to the restitution of harmonic or quasi-harmonic
signals, in stationary or quasi-stationary state,. It performs a
local restitution of desired signal in temporal space, at each
time step, using all together the different differentials measured
around the processed time. Because signals are processed in the
temporal space, the sampling of the measurements may be
irregular or have some gaps. It is not necessary to observe all
the signal : it is a local integration and hence it may be done in
real time.
The processing will be detailed in order to show first, the
measurements method by colocalisation and sub pixel level
image matching, second, the algorithm which performs the line
by line synthesis of all the measurements of each retinas
couples correlations, allowed thanks to the supposed quasi
parallelism of couple arrays, third, the integration step. We will
detailed also some further post-processes to this microvibrations
restitution like filtering, plugging gaps and completion on the
signal edges, eventually with correlation results between distant
non-parallel retinas. These post-processes allow us to retrieve
the most accurate signal over a maximum of time samples
without extrapolation.
2. DATA AND BACKGROUND
This article is focused on the process of only one product of a
pushbroom satellite. The available measurements are the results
of the comparison between at least two images of the same
product on the same landscape, viewed at different times with
slightly different angles. In the following, we denote by
“couple” one ensemble of 2 images leading to one time-
dependant measurement of a differential of the disruptive signal.
The comparison between the 2 images of a couple is made by
correlation (similarity) taking into account the two images
geometrical Models: Position of the satellite (Orbito), Line of
sight Attitude (AOCS restitution Loop), detector viewing
directions(DV) and terrain relief (Sensitivity depends on the B /
H ratio). Calculated residues dc (pixel displacement in the
swath) and dl (pixel displacement in time) from one image to
another correspond to deviations from the assumptions of
matching (“predicted shifts”).
These differences have several origins:
• Attitude inaccuracy: lack of knowledge and instability
• Disregard Land (MNT / MNS) (effect of B / H ratio
and the localization error)
• viewing directions model error
• correlation noise
Figure 1. 2 images of the same product with disruptive
vibrations
In the following, we state a very low B / H ratio (and thus an
insensitivity to DTM) and a good accuracy of the geometric
models returned. Taking account of the focal plane rigidity, we
can thus make a physical interpretation of dc(t) displacements
as a sensor roll differential and dl(t) displacements as a sensor
pitch differential. But measurements by correlation are very
noisy. It is a major error contributor that must be taken into
account in algorithm design, indeed, standard deviation of
correlation noise (supposedly Gaussian) may range from. 10%
to 20% of dc or dl.
In 2002-2004, on SPOT5, the correction of absorbed rocking
of the line of sight corrector mirror on the image have been
studied, because the tranquilisation time was chosen too short
at the beginning of the in-flight commissioning.
First, A. Bouillon has analyzed and calibrated this instability
phenomena by correlation between Pan HMA retina and HMB
retina (Breton, 2003). She found a correction model
corresponding to an extinguishing Sinus with a linear
frequency drift. Secondly, J. Jouvray has created an image
processing correction (for CNES during a training period in
2004) which computes correlation between PAN / XS with
prediction of localisation (B / H ratio ~ 0,017) and computes
the correction by a global synthesis with least squares (on all
points of measurement), modelising each shift with simple
analytical formula of partial differential.
The geometric models has been then refined with 20
parameters: vibration roll with absorption, amplitude and chirp
characteristic, attitude LF (polynomial degree 3),
magnification and LF DTM errors.
PAN XS(1 2 3.4)
Défilement
; \ Sat
capteurs PAN et XS
Visé« Visée Visé«
oblique vertical« oblique
Figure 2. SPOT5 characteristics
28