i| 2004 
?"attern 
pursuit 
, 1985. 
cience, 
pattern 
pp. 55- 
1ension 
cations. 
pectral 
nent of 
versity, 
pattern 
is and 
nsional 
School 
y, West 
al and 
Modern 
high- 
ce and 
ing and 
signal 
Pattern 
1 time- 
41(12), 
cessing. 
jn from 
; Trans. 
|-88. 
ases. In: 
of high 
. thesis, 
Purdue 
DISCHARGE ESTIMATION 
FOR THE INVISIBLE AMAZON RIVER BRANCHES IN SAR IMAGES 
Kyoichiro Katabira *, Susumu Ogawa °, Takako Sakurai ¢, Mikio Takagi ¢ 
" Center for Spatial Information Science, University of Tokyo, 4-6-1 Komaba, Meguro-ku, Tokyo, 153-8505, Japan 
katabira@iis.u-tokyo.ac.jp 
? Faculty of Geo-Environment Science, Rissho University, 1700 Magechi, Kumagaya, Saitama, 360-0194, Japan 
ogawa@ris.ac.jp 
* Department of Applied Electronics, Science University of Tokyo - takako@rs.iis.u-tokyo.ac.jp 
“Graduate School of Engineering, Shibaura Institute of Technology - mikio@sic.shibaura-it.ac.jp 
KEY WORDS: Forestry, Geomorphology, Monitoring, Visualization, SAR 
ABSTRACT: 
Recently, desertification and degradation of water cycle by deforestation in the Amazon, South America, have become a serious 
problem. In this study, it was attempted to estimate the discharge of the invisible Amazon River branches in the JERS-1/SAR images 
for the purposes of the periodical environment and disaster monitoring. First, by SFP filters for SAR images, some traces of the very 
narrow open-water channels were visualized, which were not found in the original SAR images because of the resolution with less 
than 20 m. Next, the river shapes were transformed into one-dimensional signals, and the spatial frequencies were calculated with 
the Fourier and wavelet analyses. Furthermore, with river geomorphology or the Hack's and Horton's rules, the characteristics of the 
river shapes such as the meandering wavelength, the amplitude, the branch length and the number of branches were extracted from 
the SAR images. Then, It was compared that the characteristics of the river shapes with the existing discharge data and derived some 
regression equations. Finally, the discharge of the Amazon River branches was estimated from the SAR images. 
1. INTRODUCTION 
Recently, desertification and degradation of water cycle by 
deforestation in the Amazon, South America, have become a 
serious problem. In this study, we attempted to estimate the 
discharge of the Amazon River branches from JERS-1/SAR 
images, which are independent of the weather. According to the 
river geomorphology, the river shapes such as the amplitude 
and the meandering are related with its discharges. For example, 
a small meandering wavelength shows low discharge, while a 
big one shows high discharge. First, we visualized some traces 
of the Amazon River branches by filtering, which were not 
found in the original SAR images. Next, we transformed river 
shapes into a one-dimensional signal, and calculated the river 
characteristics or the spatial frequencies with the Fourier and 
wavelet analysis. These characteristics mean the average of the 
spatial frequencies of the river shapes, the angle of the 
meandering, the number of branches, the river slope, the river 
length, and the drainage area. Then, we related the 
characteristics of the river shapes with the existing discharge 
data and derived some regression equations. Finally, we 
estimated the discharge of the Amazon River branches from the 
SAR images. 
2. METHOD 
2.1 Study Area 
We selected 11 stations in the Amazon Basin, and 12 scenes of 
JERS-1/SAR, which observed mostly rainy seasons from 1993 
to 1997 at the near sites of the discharge measurement stations. 
Beside, we used the discharge data of the Amazon River 
branches, Rio Madeira, Rio Jipanara, Rio Purus, Rio Guapore, 
Rio Tabajara, and Rio Juruena, which were observed from 1965 
to 1997. In the absence of discharge data on the same date of 
SAR observation, we used the mean discharge during the 
observation periods. 
2.2 Approach to Estimate the River Discharges 
First, to remove speckle noises in the original SAR images, we 
used a SFP filter and an enhanced SFP filter (1), and integrated 
two images. These filters could remove speckle noises while the 
small features were preserved. 
Next, we emphasized the pixel value differences of the 
surroundings in order to extract the very thin rivers from the 
bright characteristic spots. The very narrow open-water 
channels did not show the dark characteristics, but 
intermittently showed the bright characteristics with the corner 
reflected effect. We prepared a 5-by-5 window, and calculated 
the sum of the absolute differences between a center pixel and 
its surrounding pixel values (2). Then, we emphasized the 
isolated string scatters, and visualized very thin rivers, which 
were not found in the original SAR images (Fig. 1). 
m n 
WZ 
e — k=~m l=-n 
Xe = m n (1) 
2. 2" 
kz-m l=—n 
|. if Iz. - z,| «2o, z, 
Wy = : 
0 otherwise 
TO, = 0.2536 
m 
= a 2 
x, = > Sis. zy (2) 
k=—m [=—n 
889.