ISPRS Workshop on Service and Application of Spatial Data Infrastructure, XXXVI (4/W6), Oct.14-16, Hangzhou, China 
landscape distribution in this area using the knowledge based 
classification model. 
2.2 Data Source 
Since the year 1999 we have finished several field surveys 
(1999, 2003, and2004) and collected over 500 sample points. 
Also we got field data from Third National Giant Panda Survey 
containing about 1500 sample points. In this study we selected 
750 points in the study area as the reference data and all the 
sample points took the error of no more than 10m (Figure 1). 
According to the field survey, 9 types of landscape were 
specified: conifer forest (CF), mixed broadleaf and conifer 
forest (MBC), broadleaf forest (BF), bamboo (BAM), 
shrub/grass/herb (SGH), farmlands (FAR) settlements (SET), 
water (WA), rock and bare land (RB). To check the result of the 
classification, in this study we use 375 points in classification 
and 375 ones for accuracy analysis. 
An ETM+ satellite image acquired on May 22 nd , 2001 
containing 7 bands (1, 2, 3, 4, 5, 7, and 8) was used as the main 
classification data in this study (Figure 1). Also we collected 
other spatial information such as NDVI distribution DEM, 
slope and aspect data, distance to the roads and rivers 
distribution in the reserve (Table 1). All the sample points and 
spatial data were integrated into the ArcGIS environment (UTM 
projection, WGS84 datum). 
Figure 1. The study area of our research. The Foping NR lies in 
the southern slope of the Qinling Mountains founded for giant 
panda conservation. We select an area of 10*10 km in the 
middle of this reserve as our study sample. The main data 
source is an ETM+ image acquired on May 22nd, 2001 (this 
map used its combination of band 3, 2, and 1). We select 750 
sample points as the reference data. 
2.3 C5.0 Decision Tree 
The decision tree (DT) learning model was more and more often 
used in RS classification these years (Huang and John, 1997; 
Eric et. al., 2003; Liu et. al., 2005). The advantages that DTs 
offer include an ability to handle data measured on different 
scales, no assumptions concerning the frequency distributions 
of the data in each of the classes, flexibility, and ability to 
handle non-linear relationships between features and classes 
(Friedl and Brodley, 1997). DTs could be trained quickly, and 
are rapid in execution. Besides, according to the DT learning 
model, knowledge could be realized with high accuracy (Shi, 
2002). 
Data 
Description 
Scale/Precision 
Sample 
points 
Gathered from the study since 
2003 
10m 
RS 
image 
ETM+ (band 1,2, 3,4, 5,7, 
and 8) Acquired on May 22 nd , 
2001 
28.5m (band 1, 
2, 3, 4, 5, and 
7); 14.25m 
( band 8) 
NDVI 
Derived from the RS image 
(TM4-TM3)/(TM4+TM3) 
28.5m 
DEM 
Digitized form the paper map 
(1:50000) 
25m 
Aspect 
Calculated from DEM 
25m 
Slope 
Calculated from DEM 
25m 
GIS data 
Distance to roads and rivers 
distribution in the reserve 
10m 
Table 1. The data source used in this research. We collected 
sample points as reference information. An ETM+ image and 
the calculated NDVI were used as the main data set. Also we 
acquired the DEM, aspect and slope data, road and river data to 
construct a classification data set. 
DT uses a multi-stage or sequential approach to the problem of 
label assignment. Sets of decision sequences form the branches 
of the DT, with tests being applied at the nodes. The leaves (or 
branch termini) represent class labels (Figure 2). In this study a 
See5 DT model based on C5.0 algorithm was used to acquire 
the knowledge from the data sets. The C5.0 algorithm is a kind 
of univariate DT improved from the ID3 algorithm, which 
selects the branch feature according to the decrease rate of the 
information uncertainty calculated by equation 1 (Quinlan, 
1993): 
H(X /a) = p(a = dj)p(C i /a = aj) log p(C, /a = a,) (1 ) 
i i 
Where a = the value of one feature 
C = the class label 
H (X/a) = information uncertainty of feature a 
The feature with the minimal H(X/a) will be selected as the 
branch one. 
2.4 Rule base knowledge classification 
The rule-type of knowledge could be used for classification 
more effectively than the tree-type. So the classification tree 
would be converted into rules after finished DT learning in See5. 
In this study the knowledge engineer in ERDAS 8.7 
environment was used to build the knowledge base. In this 
engineer, all the classes will be treated as hypothesises and will 
be concluded from several conditions of variables according to 
the rules we got from the DT (Figure. 3).