nbul 2004 
spread 
RS: 
RS: 
ny (a «€ b) 
rplane,;); 
1; loser: — 
loser in 
to include 
unction 
im in the 
n, we will 
network, 
da, 2001- 
TION 
s (ETM+) 
scanning 
urface via 
7, 2000). 
NIR), the 
i) regions 
chromatic 
nges, and 
bits. The 
els, which 
h 3 bands 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
  
TMI, TM4, TM6 in blue, green and red, respectively. We will 
use these three bands as the inputs to the RBFC and BPNN 
because they represent most discrimination among classes. This 
site mainly contains eight types of land cover, which are water, 
urban imperious, irrigated vegetation, barren, bosque, 
shrubland, natural grassland and juniper. In the urban area, 
buildings are blocked by streets and sometimes covered with 
vegetation. So the degree of class mixture in the urban area is 
high. Besides, shrubland, natural grassland and juniper are 
highly mixed too. These bring up the difficulty in the land 
cover classification. For comparison both RBFC and BPNN are 
used to generate the land cover map of Figure 1. 
  
  
Band Number Spectral Range Ground Resolution 
(um) (m) 
TMI (Vis-Blue) 0.450 - 0.515 30 
TM2 (Vis-Green) 0.525 - 0.605 30 
TM3 (Vis-Red) 0.630 - 0.690 30 
TMA (NIR) 0.750 - 0.900 30 
TMS (Mid-IR) 1.550 - 1.750 30 
TM6 (TIR) 10.40 - 12.50 60 
TM7 (Mid-IR) 2.090 - 2.350 30 
TM8 (Pan) 0.520 - 0.900 15 
  
Table 1. Landsat 7 ETM+ bands, spectral ranges, and ground 
resolutions 
  
Figure 1. ETM+ image with 3 bandsTM 1, TM4, TM6 
displayed in blue, green, and red, respectively 
3.2 BPNN classification 
There are three input nodes, eight output nodes and one hidden 
layer with ten nodes for BPNN. First, each band is normalized 
to be in [0, 1]. The output node representing a class is defined 
as 1 (unity) if the input data point belongs to the class, 
otherwise 0. The training of BPNN adaptively adjusts the 
learning rate and the momentum [7]. It runs 10000 epochs and 
all the training data are computed once for one epoch, so each 
data point is passed 10000 times. The classification result is 
shown in Figure 2. As shown in Table 2, the classification 
accuracy of urban impervious, shrubland, natural grassland and 
juniper are low because the mixed pixels in these classes are 
not distinguished well. 
19 
  
  
  
ENS Water (WT) A Urban Impervious (UI) 
Ls] Irrigated Vegetation (IV) [1 Barren (BR) 
8 Bosque (BQ) n J Shrubland (SB) 
[ — ] Natural Grassland (NG) BB Juniper (JP) 
    
Figure 2. BPNN classification result 
3.3 RBFC Classification 
The input and output nodes of RBFC are defined the same as 
those of BPNN. The number of nodes in RBFC is determined 
in the training process and it uses 1000 nodes (or rules) to adapt 
to the training data. Each data point is only passed once in the 
training so it reduces the training time a lot. The order of 
training data is first randomized to reduce the training error. 
The classification result is shown in Figure 3 and the accuracy 
matrix of RBFC is shown in Table 3. The classification 
accuracy of urban impervious is improved from 88.84 percent 
to 95.31 percent. At the same time, the classification accuracy 
of shurbland is reduced from 86.51 percent to 71.43 percent 
and the classification accuracy of barren is reduced from 97.29 
percent to 94.86 percent. However, the classification accuracy 
of natural grassland is increased from 28.36 percent to 31.34 
percent and the classification accuracy of juniper is increased 
substantially from 36 percent to 55.43 percent. So the overall 
accuracy is slightly increased a little from 88.46 percent to 
88.64 percent. 
One of the rules in RBFC is: IF s; is N (0.93, 0.19), s> is N 
(0.95, 0.19) and s; is N (0.58, 0.19) THEN y, = 0.00058s, + 
0.000425; - 0.000195; -- 0.00025,..., yo 0.002805, - 0.005565; 
+ 0.00678s; + 0.00730. The input space is distributed with 
1000 clusters and there are more rules where the outputs 
change greatly. The outputs (O;,.., O,) from RBFC are 
satisfied with 0« O; «1 and S 0. = | after the RBFC is trained 
izl 
well. RBFC gives the fuzzy classification result, and it is 
constructed by interpretable fuzzy rules. 
4. CONCLUSION 
A radial basis function based clustering method is used in 
mulispectral image land cover classification. It improves the 
training process tremendously because the training data is 
passed once to the algorithm. Its effectiveness is demonstrated 
by the study of Landsat 7 ETM+ image classification. It can