29 
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part Bl. Beijing 2008 
Figure 3 shows the results of this method: the 2 images before 
correction and after and another image of the same landscape 
at an other day as reference. The results led to a good 
correction, allowing nice-looking images but not sufficiently 
accurate: with a sought signal amplitude of 3pixels , it 
remained a 0.1 pixel residual RMS error. 
Figure 3. Image with instabilty on the left, corrected (middle) 
and another image without vibration 
Moreover, this method based on a single image couple is 
inadequate for the following cases: when the disturbances 
model is unstable, when clouds are widely present, when there 
are inaccuracy on partial differential calculation and remaining 
ambiguities between DTM / Pitch / Line of sight error. 
This results showed however that a similar process can be 
efficient using several differentials, low B/H ratio,another 
integration method, and without VLF errors correction but only 
MF, HF characterisation. That effort was continued in a 
research and development work with Astrium in 2003 and then 
a processing unit had been built by the CNES, which is 
presented in this paper. 
3. METHODOLOGY 
3.1 Overall process 
So the main issues related to attitude improvement on image 
processing are: 
• to measure the geometric significant disruptive 
signal frequencies, 
• to separate disruptive microvibrations effects from 
MNT effects and from correlation noise and errors, 
• to calculate the absolute signal correction using 
several differentials inputs without using too much 
analytical modeling hypothesis 
• to apply quickly and automatically the corrections 
on an image. 
To reduce the measurements noise and errors, we avoided false 
correlation using a a-priori criterion, we then post-filtered 
results by synthesising them on each line of each image couple 
assuming a parallelism of retinas, The vibration is then 
obtained by synthesising at time t the attitude differentials of 
the various couples in relation to a "not rigid" vibrations 
model.. This need of filtering through a model is paramount, 
this model depends on the satellite and its disturbances. 
However, these data pertain only geometric aspects observable 
on images. First, microvibration frequencies above the 
correlation cut-off sampling frequency can never be corrected 
(Shannon theorem) and aliasing effects are possible. Second, 
because the method is based on expected (calculated) shifts 
between two images, the bias, drift, and VLF signals 
perturbing the system flight cannot be retrieved, as we would 
not be able to estimate if these observed signals, for example 
estimated on very long record length, are real or are coming 
from the measurements method itself. Third, if the differentials 
are not independent some frequencies are blind or heavily 
depreciated. 
The processing is generic and will be conducted in 4 stages as 
shown in Figure 4. 
Couple 1 Couple 2 Couple n 
Figure 4. Algorithm flow for instability correction 
The first step computes co localisations prediction (with or 
without MNT) and correlation masks to focus on the relevant 
points, then measures shift between images by correlations. It 
takes advantage of the various satellite retinas to increase the 
information at the same moment: it has several pairs of 
correlations. This first step gives, for each couple,, n line-shift 
and n column-shift temporally sampled on each correlation line. 
The following steps correspond to the heart of the system: the 
interpretation of these shift measurements and their 
transformation into attitude correction signals. First, the 
instant synthesis works on the shifts in the correlation line and 
obtains a roll and pitch differential AR(t, xj) ) & AT(t,xj) in 
assumption that the two retina of couple are almost 
parallel and separate of xj. 
Second, the temporal integration of the attitude differentials 
estimates the absolute value of the disturbance at each 
correlation line. In the overall process, it is defined as a plug-in 
which depends on the satellite and available disturbances. 
Several methods are possible, they must be robust to inputs 
gaps and noise. A local innovative method has been developed 
with assumptions that the vibrations are PHR system-like: 
quasi harmonic signals and almost stationary with 
homogeneity of differential (same convolution, spatial 
orientation...). 
An iteration is possible in order to improve the accuracy of the 
results, taking into account the estimated vibration to refine the 
instant synthesis and compute again the correction signal.