The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B4. Beijing 2008 
381 
Where 
TRRI Value of index 
Ij Digital count of channel i 
n Number of spectral channels 
I m ax Maximal digital count for given quantization level 
A Spectral channel difference 
With thin cloud, it is more difficult. It is usually mixed by some 
dry objects like dry sand or dry barren. As we know, in 
wavelength from visual to infrared, reflectance of dry barren is 
always going up according to growing of wavelength. And 
reflectance of water is always going down. Cloud includes 
water. So, reflectance curve of cloud will be going down and it 
will be low in the last band of ALOS-AVNIR2 image. The 
difference of thin cloud reflectance and dry soil reflectance is 
shown in the below chart (pixel values are gotten from Landsat 
TM image). 
Figure 6: Cloud gotten by this method (Black) 
Level 1 Level 2 
Non-Cloud 
Mixture 
Cloud 
TRRI 
—► 
Figure 5: Model for defining cloud 
To define cloud, TRRI index is used to divide image to 3 parts: 
thick cloud, non-cloud and mixture by 2 level values (as in 
above figure). In the mixture part, reflectance of thin cloud in 
blue band (band 1) is higher than that of dry soil and reflectance 
of thin cloud in band 4 is lower than that of dry soil. So, thin 
cloud and other dry objects can be separated by Cloud -Soil 
Index (CSI). This index is calculated as following: 
CSI = 
Bandi - В and 4 
Bandì + В and A 
Integrating TRRI index and CSI index can separate cloud and 
non-cloud quite well. This model can apply for every type of 
multi-spectral optical image. 
3.2 Refining cloud 
Cloud after defining still has some mistakes. Single pixels exist 
in some places. These pixels were estimated by visually that are 
some dry objects like buildings and something like that. To 
correct these mistakes, a program was developed to remove 
single pixels. After defining cloud, if result has some single 
pixels, this program can be used to make a better result. Result 
is shown in the following figure: 
Figure 7: Before remove single pixel (left) and after remove 
single pixel (right) 
The around cloud pixels are mixture of cloud and other objects. 
So, that is very difficult to define. In this study, the around 
cloud pixels are extended from cloud. An independent program 
was developed for this purpose. By this program, the cloud after 
defining can be extended to cover all real cloud. Example for 
this function is shown in the figure 9. 
3.3 Getting shadow 
Result of cloud definition is shown in the figure 6. In this result, 
boundary of cloud is still white. That is mixture of cloud and 
other objects, so that is difficult to define. This problem will be 
solved in the section Refining Cloud. 
Shadow is always difficult problem. Shadow of cloud on the 
different objects will have different values. So it is very 
difficult to define. In this study, shadow is interpolated from 
cloud. For each image, average distance from cloud to its 
shadow is determined. Direction of shadow is also estimated. 
Based on this information, a program was developed to 
interpolate shadow from cloud. Result is as following (black is 
cloud and dark-yellow is shadow):