171 
Figure 1 D edaon tree model for land cover m apping 
In their simplest form, decision tree classifiers successively 
partition the input training data into more and more 
homogeneous subsets by producing optimal rules or decisions, 
also called nodes, which maximize the information gained and 
thus minimize the error rates in the branches of the tree 
(Safavian & Landgrebe, 1991; Weiss & Kulikowski, 1991). 
techniques with the data derived from the mineral resources 
regulatory agency. 
2.2.1 Water interpretation and classification 
Before the interpretation, we first make the spectrum model 
curves of all main land covers (wood field, plantation, bare 
field, paddy field, water body, mine, residential area).After 
much comparative analysis, we find out the 
(TM2+TM3)-(TM4+TM5)>0 in paddy field, water body(Fig.2, 
Fig.3), then we utilize the0<TM3<60 and 0<TM5<40 to 
eliminate the mine and residential area by error. Considering 
that the TM4 is distinct in paddy field, water body, we utilize 
the condition TM4>40 to eliminate the paddy field by error. So 
far, we have got the exact water location map. Before the 
classification, we first select three kinds of typical water model 
including the seriously contaminated water, mildly 
contaminated water, uncontaminated water. According to the 
spectrum model curve of the TM images (all the seven bands), 
we find out the difference is discriminative in the value of the 
band three (Fig.4, Fig.5). After many tests, the determinant 
criteria are proved to be reliable: serious contaminated water 
(TM3 (40-60)), mild contaminated water (TM3 (30-40)), 
uncontaminated water (TM3 (<30)). Based on the above 
analysis, we get the water quality and location map in Daye 
area. 
2.2.2 Vegetation interpretation and classification 
Figure 2 Spectral profile of water for TM1986 
Figure 3 Spectral character profile of water for TM1986 
Figure 4 Further spectral profile of water for TM1986 
Figure 5 Classfication spectral profile of water for TM1986 
The use of decision trees as a viable alternative to more 
traditional classifiers in remote sensing research has been 
explored primarily within the context of global or continental 
scale land cover classifications (DeFries et al., 1998; Friedl & 
Brodley, 1997; Friedl et al., 2002; Hansen et al., 1996, 2000; 
Strahler et al., 1999). 
In this study, we aim to capitalize on many of the ‘lessons 
learned’ using decision trees for coarse-scale land cover 
applications.We will examine the performance of these 
According to the field investigation, the relationship among the 
Normalized Difference Vegetation Index (NDVI), vegetation 
types and the elevation is analyzed synthetically. Based on the 
result of the water is masked, we set up the decision criteria of 
the vegetation distributing in Daye area: wood field 
(NDVI>0.45, DEM>100); shrub (0.3<NDVI<=0.45, DEM>50), 
we get the vegetation distributing map in Daye area.