International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
b. Spatial feature 
Spatial features are the shape, size and the edge of 
the target. 
c. Texture feature 
Textures provide important characteristics for the 
analysis of many types of images including natural 
sensing data and biomedical modalities. 
In this study, we mainly used the NDVI and TC 
feature images, these feature images and the 
preprocessed TM image were input into the C5.0 
classification platform together. 
3) Extract samples and create sample database 
A sufficient number of training samples and their 
representativeness are critical for image classifications 
(Hubert-Moy  etal.2001, Chen and Stow 2002, 
Landgrebe 2003, Mather 2004). Training samples are 
usually collected from fieldwork, or from fine spatial 
resolution aerial photographs and satellite images. To 
grantee the precise of the training samples, we did 
some fieldwork. The information of the sample area 
is: 
Location: southeast coast of Australia 
Latitude and longitude: 33°03'00"~34°47'00"S, 
149°23'00"~152°01'00"E 
Climate: wet climate 
Ecoregions: Southeast Australia temperate forests, 
and Eastern Australian temperate forests 
Some of the training samples were collected from 
the LANDSAT TM images. This process can 
implement on remote sensing or GIS softwares. The 
features displaying on the image are inflected by 
many factors, such as climate, terrain. According to 
ecoregions and months, the 18 scenes images were 
dm 
LeftiMLC result) 
middle(classified image) right{C5.0 result) 
divided into three teams. We chose three or four 
scenes images to select samples, and guaranteed every 
scene's sample points were no less than 500 points. 
4) The creation of the classification rules 
When we have sufficient training samples and 
good feature files, the next procedure is to get 
classification rules. Here, the C5.0 classification 
platform can be adopted. According to the grouping of 
the above procedure this study can get three decision 
rules. 
5) Classification 
The last procedure is to use the decision rule to 
classify. In the study, every scene of the 18 scenes 
images of Victoria used one of three rules to 
implement classification. When every scene image 
was classified, the classification results were 
mosaiced, and then we got the classification result of 
Victoria. 
3 EXPERIMENTAL RESULTS AND 
ACCURACY ASSESSMENT 
3.1 Experimental Results 
Maximum likelihood classifiers are frequently 
available and widely used for land cover classification 
from multispectral imagery. In the study MLC 
classification was also used to classify the images of 
Victoria. Also, the same training data was used to 
classify one scene by using C5.0 classification method 
and MLC. 
   
  
   
   
  
    
WE unelaisified 
Bl cropland 
Bl forest 
Li graxs 
| shrub 
water 
wetland 
Figure 4. Some comparisons of MLC result and C5.0 result 
Figure 4 shows some visual comparisons of MLC 
result and C5.0 result, we can easily found that the 
C5.0 result is closer to the actual classified image. 
Figure 5 shows the classification result of Victoria.