International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
[tert Training samples Verifying samples 
] 2 3 A 5 6 7 8 9 10 
Input band-1 1.212 | -0.736 | -0.885 | -0.587 | -1.290 | -0.141 1.672 1.131 | -0.317 | -0.060 
band-2 1.153 | -0.628 | -0.824 | -0.677 | -1.333 | -0.176 1.637 1.235 | -0.357 | -0.029 
band-3 1.082 | -0.313 | -0.736 | -0.684 | -1.082 | -0.456 1.825 1.258 | -0.574 | -0.319 
band-4 0.476 | 1.606 | -0.714 | -0.654 | -1.071 | -0.595 1.546 0.714 | -0.654 | -0.654 
band-5 -0.887| 2.070 | 0.222 | 0.222 | -0.795 | -0.979 | 0.961 -0.979 | 0.499 | -0.333 
Object SS 2005 |..3.65 15.65] 19.65 | 20.65 |. 25.15 35 24 IS 15 | 21.55 
CODmn 1.7 2.05 23 1.85 2 2 1:95 r7 1.8 1.9 
DO 53 4.76 5.68 5:15 5.8 6.5 5.95 4.78 6.2 6.38 
T-P 0.055 | 0.03 0.04 | 0.035 0.03 0.03 0.055 0.05 0.04 0.025 
T-N 0:855|,0:605 | 0.918 | 0.715 | 0.855 | 0.765 0.655 0.79 0.79 0.875 
chl-a/lug.L' | 5 5 5 4 3 5 4 5 6 5 
Table2. Original data of the artificial neural network model/mg.L" 
In the training process, size of s, which was amounts of neural and output parameter dimensions by the way of adjusting 
cell in eryptic layer, was justified gradually. Ultimately s was 
confirmed as 10 by analyzing accuracy of verifying samples, 
training time and iterative times. 
The verifying results were showed in Table 3 after testifying 
samples were inputted to the having been trained ANN model. 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Water quality SS (COD, DO | T-P | TN chl-a 
parameters® * ug.L 
2.09*4 1.66*4 5.23** 5.6841 843 | 4.91 
Inversing | [^5 29k 1.64*1 4.98*] 3.5941 621 | 4.11 
results * 
26.78** 2.01*4 6.52** 3.14*1 6.69 | 4.88 
Synchronous | 2:40*] 1.70*1 4.78*f 5.00 7.90 | 5.00 
monitoring | 15.15** 1.80*4 6.20*? 4.00*4 7.90 | 6.00 
fesultss «T e 1901 638 250-1 $75 | 5.00 
-13.014*-2.15*» 9.33* 1 13.52 6.69 | -1.84 
Inversing  [33.95-h-8.81+h-19.671+10.264421 44/3146 
errors®® 
24.29%} 5.79*4 2.26* 25.63**23.50| -2.35 
Table3. The results of inversion and validation of the 
model/mg.L” 
It could be concluded that inversing errors could be controlled 
below 25%, except for SS and chl-a in second verifying data, 
which exceeded 30%. The results were satisfactory. 
4.2.2 Result Analysis 
Results indicated that the artificial neural network was well able 
to inverse water quality parameters from remote sensing image: 
(1) Inversing effect of the model was good (Table 3). Inversing 
errors could be controlled less than 25%, except for SS and chl- 
a in second verifying data. If some measures as follows were 
taken in data collecting, processing and analyzing, inversing 
precision could be improved further. 
(2) Ability of the model in simulating complicated relations was 
strong. ANN could realize nonlinear mapping between input 
680 
weights and biases of neural nodes (Cong, 1998). Results 
indicated that the ANN model had good simulating effects. 
(3) The model inversed multi-water quality parameters 
simultaneously. The paper gave a model that SS, CODMn, DO, 
TP, TN and chl-a could be inversed from TM image using a 
trained ANN in the same time. 
Reasons of one or two parameters having higher errors might be: 
synchronous experiment's internal errors, for example, boat 
stirring in shallow water or laboratory error; little sampling 
points were placed, which might result in water quality 
parameters bad-proportioned distribution in their concentration 
ranges; processing errors of remote sensing data, especially 
atmospheric correction errors; structure errors of the model. 
The research showed that the first item error could be avoided 
or decreased by some steps, such as sampling after a moment of 
boat stopping, or increasing number of samples for each point; 
the second item error could be removed through steps of 
increasing sampling points or making these points reasonable 
arrangement; the third item error was difficult to remove. If it 
was possible, some optical experiments should be performed at 
the same time of satellite synchronous monitoring in order to 
decrease atmospheric correction errors. The last item error was 
inevitable, but adjusting neural network structure or comparing 
various amounts of neural cells could reduce the error. 
5. CONCLUSIONS 
Literature review indicated that water color remote sensing 
inversion often adopted empirical model for limits of water 
color remote sensor technology and atmospheric correction 
arithmetic. 
This paper analyzed traditional regression model through 
combining different bands and operations. The relative 
coefficients between remote sensing data and water quality data 
was not good and could not satisfy the requirements of 
application. Analysis indicated that it was resulted from the 
interaction of many water quality components, such as SS, 
chlorophyll and yellow substances. 
Inter 
—— 
Both 
the / 
resez 
com] 
On t 
BP 
conc 
invel 
acce| 
And 
appli 
this 
sync 
succi 
wate 
Bric: 
scatti 
Opt., 
Cam 
estim 
conti 
Claud 
disk 
Land 
pp. 1 
Chen 
chlor 
data. 
176(i 
Chen 
Wate 
Tsing 
Cong 
MAT 
china 
Ekstr 
chlor 
Sens. 
Kuan 
Study 
In: 
devel 
More 
color 
Richt 
appli 
159-1 
Richt 
Corre: