il 2004 
  
ng part 
ickage, 
of TM, 
COR2. 
oure 3. 
BILITY | 
"with 
ded by 
, 1990; 
tandard 
; "been 
angles, 
es, the 
p table. 
formed 
method 
| model 
cts the 
1g data 
les and 
pendent 
rder to 
; study 
jon was 
ig band 
R1*R2) 
sensing 
relative 
unts of 
on were 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
  
  
  
  
  
  
  
  
Atmospheric | dependent good band relative 
correction | variables combinations qpeíficien 
SS R1/In(R3) 0.292 
No Chia In(R1+R3)/In(R2) 0.51 
In(R1*R3)/In(R2) (0.572 
d R1/In(R3) 0.411 
Yes (R1+In(R2))/In(R3) 0.428 
Chl-a In(R1+R3)/In(R2) 0.64 
  
  
  
  
  
  
Tablel. Results of Regression analysis 
3.2 Analysis 
Table 1 showed that: 
€ Most combinations had bad relative coefficients between 
concentration of water quality component and remote 
sensing. data. 
€ As Table l, only six combinations were extracted from 
the total 356, but their relative coefficient were still not 
good. 
€ Relative coefficient could be improved a little after 
atmospheric correction. 
€ It was hard to build an applicable regression model for 
quantitative analysis of Case II water body. 
The main reason was: 
€ Water components interacted each other strongly. 
Relations between image bands and water components 
were crossed and non-corresponding relative. 
€ The wider band width also decreased their relativities. 
4. THE ARTIFICIAL NEURAL NETWORK 
INVERSING MODEL 
Theoretical research has proved that if an ANN includes biases, 
at least a S-style cryptic layer and a linear output layer, it can 
approach any rational function (Cong, 1998). And theoretical 
research of water color remote sensing indicates that remote 
sensing data are correlated with water body components and 
their concentrations directly after atmospheric effect is 
removed . However, because water body ingredients affect each 
other, traditional inversing method, such as building a relative 
function between band data and components cannot solve it. 
The artificial neural network is suitable to simulate such 
complicated relationship. Therefore, theoretically, it is feasible 
to use ANN in water color remote sensing inversing research. 
4.1 Model Structure 
Figure 4 showed a concept chart of artificial neural network 
structure used in this study. Every input node represented a TM 
band, and these input data were distributed to each node of 
cryptic layer to operate. Output values of cryptic layer were 
inputted into output layer and operated again. The output values 
of output layer were parameters interested by users. 
  
t Kuang, C., 1999. Identification of Water Color Remote 
Sensing Theoretical Model and Several Aspects in Its 
Solution. Tsinghua University doctoral dissertations, pp. 
15~86. 
679 
  
Figure 4. The structure of artificial neural network 
In detail, the paper adopted a two layer BP artificial neural 
network. 
p; (i-1* r), r-5, represented five inputs which were bandl to 
band5 of TM image data; 
IW, , represented the weight between inputs and neural cells of 
first layer. lw, represented the weight between outputs of 
first layer and neural cells of second layer. 
2 
b and b^ were biases of neural cells of first and second layer 
respectively. 
f^ f was transfer functions. f! often adopted S-style 
transfer function, in this paper the model used hyperbolic 
2 
tangent S-style transfer function. f adopted linear transfer 
function. Theoretical research has proved that an ANN, which 
has biases and at least a S-style cryptic layer and a linear output 
layer, can approach any rational function (Cong, 1998); 
2 ^ 
a! and a? were output values of first and second layer (see also 
Figure 4): 
a - f'üW, "pb 
a mf" LW *a vb) 
m, the number of neural cells in the second layer, was 
determined by amounts of object of ANN model object. The 
paper hoped that this ANN model could simulate SS, CODy,,, 
DO, TP, TN and chl-a, so m=6; 
s, the number of neural cells in the first layer, was determined 
by neural network training process. According to results of the 
paper, s was equal to 10. 
4.2 Training Results and Its Analysis 
4.2.1 Data of Model Training and Verifying 
Generally the input data should be normalized in order to 
improve training efficiency and accuracy of neural network 
(Zhan ,2000). Table 2 gave the input data and object values.