Since differences of target size and spatial structure still exist in 
high resolution remote sensing images, it is difficult to reflect 
the rich object and spatial semantic information only in a single 
spatial scale. Therefore, we need different scales to express and 
describe different size target. The structure of image object 
obtained in different segmentation scales represents different 
scales image object information, in which a smaller object is a 
child object of a larger object (Tan, 2007). Since each object 
has interrelated spatial feature information with adjacent objects, 
child object and parent object, the purpose of our experiment is 
to explore the spatial feature of interrelated objects and add it to 
the multi-feature combination model. 
3. FEATURE DESCRIPTION 
The purpose of information extraction is to distinguish 
interested regions in remote sensing image. There are different 
methods corresponding to different targets. Especially, some 
specific thematic objectives need to fully consider the 
characteristics of data (Datcu, 2002; Anthony, 1997; Zhou, 
1999). Features description is object-oriented expression for 
latent knowledge in the primitive obtained from image 
segmentation. In addition to visual features such as spectrum, 
feature description also includes the object shape, spatial 
relationship and terrain features (Yang, 2009). 
3.1 Spectrum features 
In addition to traditional statistical value of image gray such as 
histogram, variance, mean, the normalized difference snow 
index is suitable for extracting glacier information since it is 
very sensitive to the changes of water content in snow and ice. 
In the following formula, 7x is the red band reflectance and 
Pr is the green band reflectance (Guo, 2003). 
NDSI P ge 4 7 P Gren 
Pre à + P Green 
(6) 
3.2 Shape features 
Generally, the boundary of ice cover is clear, around which 
there is often a large number of moraine. Meanwhile, some 
clear curved contours are left behind the ice tongue in the 
process of glacier retreat. In addition to the compactness and 
smoothness mentioned above, there are some shape indexes to 
describe glacier feature. 
3.3 Spatial relation features 
As for each object in the spatial relationship, we can calculate 
the mean difference among the objects, and give a weight 
according to border length or area size to achieve classification 
and clustering of object (Definiens, 2007). Some statistics can 
reflect the spatial distribution features of pixels enclosed by 
different object. 
3.4 Terrain features 
Glacier is distributed above the perennial snowline of the alpine 
region, in which valley glaciers are located in canyons of high 
mountain. Since the shapes of ice tongue, ice pillar and ice cliff 
are closely related to the terrain factor, the terrain features have 
more important role to the glacier information extraction. 
Digital elevation model (DEM) mainly describes spatial 
distribution of region landform, and can determine slope, aspect 
129 
and relief degree of earth's surface (Song, 2007), so we can 
identify and extract glacier information by the DEM. 
4. GLACIER INFORMATION EXTRACTION BASED 
ON MULTI-FEATURES COMBINATION MODEL 
In our researches, the data we used are both of 2.5 m 
panchromatic and 10 m multi-spectral SPOT-5 images. After 
pretreatment, we select a sub scene fused image of eastern 
Qinghai-Tibet Plateau as trial data, with a rectangle area of 436 
km’. The original image is shown as Figure 1 (a), and the 
enhanced image by histogram is shown as Figure 1 (b). The 
glacier information extraction of multi-feature combination 
model mainly includes the following steps: (1) multi-scale 
image segmentation; (2) the bound identification of ice and 
snow; (3) glacier information extraction. 
    
(a) Original image 
  
(b) Enhanced image 
Figure.1 Image of Qinghai-Tibetan Plateau trial area 
4.1 Multi-scale image segmentation 
In the process of multi-scale image segmentation, we select 0.1 
and 0.7 as the weight of shape and smoothness. Meanwhile, we 
set the segmentation scale as 500, 200,100 and 50 and get 
different results as Figure 2 (a) to (d), with 72, 354, 1036 and 
3290 image objects, respectively. Obviously, over-segmentation 
phenomenon exists in the Figure 2 (c) and (d), and under- 
segmentation phenomenon exists in the Figure 2 (a) and (b).