). The 
image 
th red 
is the 
of the 
| for extensive 
) (Muller and 
/M3 matchers 
et al., 2002), 
t predecessors 
ig techniques. 
nalised cross 
it can account 
timal method 
nt in imagery 
t any changes 
les and noise 
TSR images. 
tive disparity 
ontinuities. À 
ity population 
for example. 
as intensity 
  
  
changes within the imagery. The statistical measures (regional 
mean and standard deviation) used in the NCC algorithm 
assume the intensity values within the patch have a Gaussian 
distribution, at discontinuities this assumption may be 
invalidated due to the occurrence of multi-modal intensity 
distributions. This leads to erroneous disparity measurements at 
discontinuities. ~~ Here we implement a stereo matching 
algorithm based on the Census algorithm (Zabih and Woodfill, 
1994) to achieve better results at depth discontinuities. 
2.3 Census 
The Census transform belongs to the group of patch based non- 
parametric transforms, including the Rank transform (Zabih and 
Woodfill, 1994) and Ordinal measures (Baht and Nayer, 1998). 
These transforms do not rely on the pixel values themselves; 
rather they encode the ordering of the pixel values. The Census 
transform in addition to encoding the ordering of the pixels also 
stores the spatial structure of the pixels within the patch. This is 
achieved through the definition of a bit string which encodes 
the relationship of a pixel of interest to those pixels within its 
local neighbourhood, defined by the patch size. Here, we apply 
the modified census transform (MCT, Froba, 2004), which 
replaces the central pixel value with the mean of the patch, 
which further improves performance. MCT can be described 
formally as follows: 
FG)- 9.606)0) — 1n 
Where N (x)is the neighbourhood centered on the pixel X so 
that N(x)e N(x)ux. The comparison function 
C(I(x),1(y)) is 1 if I(x) « I(y), where I(x) is the patch 
mean intensity and I(y) is a pixel intensity from the local 
neighbourhood. Lastly, ® denotes a concatenation operation, 
which generates the bit string. 
As non-parametric transforms depend solely on the ordering of 
the local intensities and not the magnitudes they are robust to all 
radiometric distortions that do not change the ordering. Further 
they are less susceptible to the effects of intensity variations at 
discontinuities (Zabih and Woodfill, 1994), leading to more 
accurate disparity estimations and less smoothing across 
discontinuities. For a detailed analysis of cost metrics see 
Hirschmuller and Scharstein (2008). 
Applying MCT to the AATSR stereo image pair, stereo 
matching can then be achieved by finding the most similar 
comparison bit string from within a search window to a 
reference bit string. Similarity is defined using the Hamming 
Weight, effectively the sum of the Hamming Distance also 
referred to as an XOR operation. Comparisons are achieved 
efficiently by converting the bit strings into bit numbers and 
applying bit twiddling methods to rapidly evaluate the 
Hamming Weight. The raw output from MCT for a scene over 
Greenland is shown in Figure 1b. 
3. VALIDATION 
Validation is carried out against LiDAR cloud layer 
measurements at Skm spatial resolution obtain from the 
CALIOP instrument. For the month of June 2008 all AATSR 
scenes over-passing Greenland processed using MCT and the 
CTHs were extracted using the Mannstien camera model (Denis 
et al., 2007). From this dataset a total of 6 AATSR orbits were 
found to have collocated CALIOP measurements within their 
swath, with at most two minutes between acquisitions. Here, 
we present an initial inter-comparison using one of these 
collocated datasets for stereo results derived from the 0.55um 
channels (the nadir image is show in Figure 1a). 
The collocation methodology involves a number of steps: 
Firstly, a 5x5 pixel median filter is applied to the AATSR CTH 
result, this smoothes the AATSR data to a similar resolution to 
the CALIOP data (see Figure 1c). Secondly, the data are 
collocated using the associated lat/lon grids. Each collocated 
AATSR pixel is then compared to the collocated column of 
CALIOP cloud top (CTL) and cloud bottom layers (CBL). 
Comparison to CTL and CBL is required as CALIOP has 
increased sensitivity to cloud compared to AATSR. | AATSR 
resolves the cloud height where the cloud reaches a certain 
optical thickness, whereas CALIOP determines the cloud top to 
be where it first encounters a cloud signal. The matching 
CALIOP cloud layer from the column is the layer height which 
has the minimum distance from the AATSR CTH. Lastly 
outlier removal is performed on each collocated data set, with 
an outlier defined as any collocated measurement pair whose 
heights differ by more than two sigma from the mean height. 
4. RESULTS 
The two sets of inter-comparison results are presented here. In 
Figure 2, the inter-comparison between the AATSR CTH and 
CALIOP CTL are presented. A total of 154 inter-comparisons 
were made. The AATSR CTH measurements appear quantised 
in comparison to the CALIOP results, this is due to the pixel 
level accuracy of the AATSR measurements leading to ~lkm 
groupings. In the CALIOP transect (shown in Figure 1c) there 
appear to be two main CTL groupings, one cloud feature at 
between 4-6km and another at 8-9km. The results in Figure 2 
show that AATSR is generally underestimating CTH in 
comparison to the CALIOP CTL. The bias between the 
measurements confirms this at -2.45 km. The RMSE is 2.76 km 
and the coefficient of determination is 0.54. 
  
AATSR CTH vs. CALIOP CTL 
  
10 
9L A e = = 
$ s, 2 
e 9 ° - N og e 
= ; : : 
x 7 eus i ze. 
E sius 
0. $ ® 
o ge 
= 5t oink eut 
< 3% js 
Q e e i Pe 
s res $ 
3 2 o4 : s 
2 3 
  
  
3 a 5 $ 
AATSR CTH (KM) 
Figure 2. This Figure presents the results of the 
inter-comparison between AATSR CTH and 
CALIOP CTL for the transect presented in 
Figure lc.