el 
re 
ag 
or 
he 
International Archives of the Photogrammetry, Remote Sensing 
    
The sensor model was applied to georeference the MISR level 
1B1 product (Poli, 2003). Two areas of interest, over Germany 
and South France, were chosen. From the test over Germany a 
GCP accuracy of 173 min X, 87 m in Y and 80 m in Z was 
achieved, corresponding to 0.6, 0.3 and 0.3 pixels (ground pixel 
size: 275m). For the test over South France, RMS errors of 43 
m in X, 45 m in Y and 152 m. in Z. corresponding to 0.2, 0.2 
and 0.6 pixels, were obtained. 
The first results are promising, because the images have been 
oriented with sub-pixel accuracy. The self-calibration was 
fundamental because it allowed the estimation of the correct 
internal and external orientation parameters. In these tests, 
significant values for the principal point displacement have 
been estimated. Without self-calibration, the RMS errors in 
GCPs were larger than one pixel. 
All of our MISR CTH and CTW calculations presented in this 
study (Section 2.3) fully rely on the operational LIB2 
georectified radiance data. In the future, however, it is planned 
to use the georectification from the described in-house sensor 
model. 
2.3 Cloud-Top Height and Motion Estimation 
Determination of CTH from ATSR2, AATSR or any two views 
of MISR proceeds along the same scheme, as illustrated in 
Figure 1. First, all images were reduced to 8-bit with linear 
stretching between the minimum and maximum values. As no a- 
priori values of the cloud heights were given to the matching 
algorithm, the number of pyramid levels for the hierarchical 
matching was chosen so that the maximum possible parallax at 
the highest level was only 1-2 pixels. Three and five pyramid 
levels were used for ATSR2/AATSR and MISR, respectively. 
Every pyramid level was enhanced and radiometrically 
equalized with the Wallis filter. According to the block or filter 
size, different cloud structures could be enhanced. In general, à 
block size of about 70 pixels was chosen at the original level, 
which was then decreased up the pyramid. Points with good 
texture were selected with the Fôrstner or Harris interest 
operator in the first or second pyramid level because it is likely 
that these same points are readily detectable in the other levels. 
If a cloud mask was available (e.g. our own cloud mask for 
ATSR2, LI RCCM or L2TC cloud masks for MISR), we used it 
for thinning of the point set to cloud points only, prior to 
matching. 
The unconstrained Multi-Photo Geometrically Constrained 
(MPGC) least-squares matching (LSM) (Gruen, 1985; 
Baltsavias, 1991) was applied hierarchically, starting on the 
highest pyramid level. After each pyramid level, quality control 
with absolute tests on the LSM matching statistics was 
performed to exclude the largest blunders from further 
processing down the pyramid. The patch size was slightly 
increased from one pyramid level to the next, from 7 x 7 on the 
highest level to about 15 x 15 on the lowest level. 
After applying the MPGC LSM algorithm, the matching 
solutions were quality-controlled with absolute and relative 
tests on the matching statistics. Additionally, meteorological 
criteria can be used in the detection of large blunders, including 
minimum and maximum cloud heights, minimum and maximum 
cross-track parallaxes, which are, after division by the time 
difference, proportional to the cross-track wind speed, or 
1157 
    
and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
filtering the cloud heights with the brightness temperature 
values from the IR. channel(s) in the case of ATSR2/AATSR. 
The resulting y-parallaxes were converted into cloud-top 
heights according to Prata and Turner (1997). The zenith angles 
(e.g. Osa, and Ossa for ATSR2/AATSR) thereby had to be 
projected on the along-track plane. The height values of the 
successfully matched points were finally interpolated to the full 
resolution grid. 
The accuracy of the retrieved cloud-top heights was dependent 
on the geometric stereo configuration expressed as the base-to- 
height ratio B/H, the matching accuracy Ay,, the accuracy of the 
georectification, including the exact values of the zenith angles, 
and the along-track motion retrieval accuracy Av'. In Table | 
the B/H values and image time differences for ATSR2/AATSR 
and three different viewing angle combinations of MISR are 
listed, together with an estimation of the height error Ah given 
an along-track parallax error Ay, of 1 pixel from matching or an 
along-track motion error Av' of 5 m/s. 
For all sensors, the height error due to motion errors is very 
prominent. In contrast to stereo image pairs from scan- 
synchronized geostationary satellites, stereo image pairs from a 
single polar-orbiting satellite are never perfectly synchronous. 
There is a time delay of seconds to minutes between image 
acquisition at the different viewing angles. The resulting errors 
in stereo cloud-top height retrievals can be quite large, 
depending on the along-track cloud motion. as pointed out in 
Table 1. If more than two non-symmetric views are available, 
the along-track parallax can be separated into the amount due to 
cloud height and the amount due to cloud motion. With only 
two views, or symmetric multiple views, which is the usual 
case, the along-track cloud motion has to be corrected with data 
from an independent source. One possible source of 
independent data is geostationary satellite cloud motion 
information. In our study, three types of geostationary data from 
the two European satellites Meteosat-6 and Meteosat-7 
(Eumetsat, 2003) were used: the Meteosat-6 5-minute Rapid 
Scans during MAP, the quasi-operational Meteosat-6 10-minute 
Rapid Scans and the operational Meteosat-7 30-minute 
sequences. The launch of the first Meteosat Second Generation 
(MSG) satellite (called Meteosat-8 since its transition into 
operational mode in March 2004) in August 2002, with a 
temporal resolution of 15 minutes, now offers a further data 
source for accurate CTW retrieval in several spectral bands. 
   
   
  
    
  
  
  
  
  
   
   
    
    
B/H ratio At [s] |Ah [m] for Ay,| Ah [m] for 
= | pixel Av' = 5 m/s 
830-1430 
Table 1. Height error caused by parallax error and along-track 
motion for various B/H and time acquisition 
difference cases. 
     
  
  
  
  
  
0.7-1.2 100-130 
  
      
    
   
  
   
  
  
  
  
  
   
For cross-track wind retrieval and along-track wind correction, 
the exact time difference At between corresponding pixels in the 
forward and  nadir scans had to be calculated. For 
ATSR2/AATSR, the time difference varies significantly over