IS, Vol. XXXVIII, Part 7B 
In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
49 
VHRR data 
VHRR data 
t orient classification is 
.lgorithm does not only 
but also on pixel spatial 
channel means, standard 
The Sixteen components 
in this study. The five 
itellite images (primary 
imponents such as albedo 
bands , solar zenith and 
iperature (LST) & Sea 
Normalized Difference 
ation of nadir and cloud 
components use directly 
> on classification. The 
ected in this study. This 
on and a knowledge-free 
this method the used 
eratures of band 4 and 5 
1, whereas the Digital 
ST and NDVI that show 
i by 0.2 (Gottsche and 
h medium area and more 
y medium scale of 50 is 
in (level 1 that called 
using infrared or albedo 
l classification technique, 
relationship between cold 
rnds were also used to 
lows the multiresolution 
Figure3. Multiresolution segmentation in analysis level 
In remote sensing studies we cannot detect a specific cloud in 
a range of Digital number in visible or infrared images 
exactly. For example, the cumulonimbus clouds (Cb) in each 
region and time can are detected with various range of digital 
number. But studies show that this type of cloud is brighter 
than others in VIS and IR images or stratus clouds (St) are 
darker than others in IR images. The other types of clouds 
can were detected similarly that is showed in figure 4 but 
texture, shape and thickness of clouds are useful option for 
decision. Figure 4 is the principle of bi-spectral technique in 
this study. 
Figure4. Brightness of each type of clouds in VIS and IR 
images.(Ito ,2000) 
3. RESULTS AND DISCUSSIONS 
In this study 8 classes were identified that were included a 
non-cloud class (sea and terrain) and 7 cloud classes (Ns, Ci, 
Cb, Cu, Cg, Sc, St). After detecting of classes and scrutiny of 
features (mean, standard deviation, to super object, shape 
(area and density) and texture (homogeneity, contrast and 
entropy), classification was carried out. The nearest neighbor 
classification of level 1 object was performed for mean and 
standard deviation of AVHRR channels of 1, 2 and 4. The 
relationships between objects were included in the 
hierarchical classification for getting the better results. The 
classifications of clouds were performed in level 2 called 
cloud level. The result of classification illustrated in figure 5. 
El ■ 
classes 
^ cirrus 
(3 cumulonimbus 
^ cumulus 
(3 cumulus congestus 
(3 nimbo stratus 
Q no douds(sea&terrain) 
HI strato cumulus 
4) stratus 
Figure 5. Classified clouds with eCognition 
Since the detected cloud in sky of each area is the dominant 
cloud on that hour and the other type of clouds may be 
existed during the day, so collecting of control points for 
accurate assessment is not possible. So, the regions that have 
the most adaptation with bi-spectral cloud classification 
theories were used as training samples optically. 
The NIR and IR channels 3, 4, and 5 of the data were 
processed for temperature and brightness. In IR image cold 
clouds are high clouds, so the colors typically highlight the 
colder regions Mid height clouds with TB below 230k were 
identified as cumulonimbus cloud. Darker clouds in IR 
images were associated to warm stratus, Strato cumulus, 
cumulus clouds and thin cirrus cloud that were colder than 
others. Using this knowledge and bi-spectral technique, the 
sampling areas were selected and error matrix and kappa 
coefficient performed using TTA MASK in eCognition 
software. The obtained kappa coefficient was equal of 0.887 
and the overall accuracy is 0.905 that illustrate the accurate of 
classification is high. The window of error matrix based on 
TTA Mask is showed in figure 6. 
Figure 6. Error matrix based on TTA Mask 
4. References 
Billa, L, Mansor,Sh, Mahmud ,A.R, Ghazali ,A,H, 2004. 
Integration of RS, GIS and MIKE 11 Hydrodynamic 
Modeling for Flood Early Warning: A case study of the