International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
registered cloudy images, and to mosaic “clean” data together. 
If the clean data pixels from different scenes are simply 
consolidated without additional processing, the final image will 
be very "speckled" and appear discontinuous. Therefore, a 
patch of pixels rather than the individual pixels is chosen to 
form the final mosaic. 
The final mosaic is composed from the images with cloud, 
cloud-shadow masks and the merged image generated from the 
merging of sub-images procedure. To suppress the visibility of 
the seam line between adjacent patches, the residual intensity 
differences between the patches are balanced using the intensity 
histograms of local patches. Secondly, the patches are made to 
overlap at their boundaries and the system will blend the image 
intensity from one patch to another in these overlapping 
regions. Finally, the images resulting from the mosaic process 
are geo-referenced to a map. The mosaic production procedure 
will put the image from the mosaic process into the map. 
3. RESULTS AND CONCLUSIONS 
An example of applying the cloud-free mosaicking algorithm 
the six cloudy SPOT panchromatic images is shown in Figure 2. 
Figure 3-(a) shows a mosaic of cloudy SPOT multispectral 
images over Singapore and the southern part of the Peninsular 
Malaysia. The resulting cloud-free and cloud shadow-free 
mosaic is shown in Figure 3-(b). The mosaicking algorithm has 
also been tested on 1-m resolution IKONOS colour images. 
In this paper, we have presented the method for producing 
cloud-free and cloud shadow-free multi-scene mosaics from 
cloudy SPOT and IKONOS images. The system has been 
implemented successfully over a large area covered by about 50 
SPOT scenes. The success of the cloud-free and cloud shadow- 
free mosaic depends on the choice of the shadow, vegetation 
and cloud intensity thresholds. Confusions arise when high- 
albedo open land surfaces or buildings are encountered. Such 
756 
confusions can be resolved by making use of size and colour 
information to classify the pixels /patches into a few broad land 
cover classes. In many cases the clouds that need to be masked 
out are much large than the individual building, an automatic 
method is developed to calculate the size of the bright patches 
in order to eliminate improper masking of these buildings. As a 
result, this procedure allows a few small cloud patches to 
remain in the mosaic. A large, very bright and white patch of 
open land surface will be considered as cloud. When the bright 
and white patch of open land does not contain cloud-shadow, it 
is still possible for this patch of open land surface to be selected 
and used in forming the final mosaic. The approximate location 
of cloud shadow can be predicted based on the knowledge of 
solar illumination direction, sensor viewing direction and cloud 
height. 
References: 
S. C. Liew, M. Li, L.K. Kwoh, P. Chen, and H. Lim, “Cloud- 
free multi-scene mosaics of SPOT images,” in Proc. 
International Geoscience and Remote Sensing Symposium, 
1998, vol. 2, pp. 1083-1085 
M. Li, S. C. Liew, L.K. Kwoh, and H. Lim, “Improved cloud- 
free multi-scene mosaics of SPOT images,” in Proc. Asian 
Conf. Remote Sensing, 1999, vol. 1, pp. 294-298 
M. Li, S. C. Liew, and L.K. Kwoh, “Generating “cloud free” 
and "cloud-shadow free” mosaic for SPOT panchromatic 
images,” in Proc. International Geoscience and Remote Sensing 
Symposium, 2002, vol. 4, pp. 2480-2482 
M. Li, S. C. Liew, and L.K. Kwoh, “Producing Cloud Free and 
Cloud-Shadow Free Mosaic from Cloudy IKONOS Images” in 
Proc. International ~~ Geoscience and Remote | ensing 
Symposium, 2003, vol. 6, pp. 3946-3948 
Inte