IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring”, Hyderabad, India,2002 
  
Hong Kong) where sediment from the land increases the digital 
count of coastal waters adequately for differentiation of turbid 
river plume water from clear sea water. The authors have been 
able to remove this noise by a simple method which employs a 
series of median filters, leaving the main boundaries in the 
image (between different water masses) essentially unchanged, 
both spatially and spectrally. 
2. METHODOLOGY 
2.1 Study area and images used 
The area imaged is the coastline of Jambi Province, Sumatra 
where, due to forest clearance, the large rivers Lagan and 
Batanghari discharge into the clear waters of the Java sea 
(Figure 2) creating distinct and extensive, turbid river plumes. 
These freshwater plumes are buoyant relative to saltwater, and 
push seawards, as a shelf, over the denser saline water, 
persisting for a distance of up to 30 kilometres from the coast. 
The Suspended Sediment (SS) within the plumes acts as a 
tracer for the seaward extent of fresh river water. The ultimate 
seaward boundary of the plumes is sharp, as are the boundaries 
of smaller eddies within the larger plumes due to differing 
levels of salinity creating non-mixing water masses. Image 
radiance generally decreases away from the coast due to the 
decay of the plumes, accompanied by lowering of the sediment 
content. 
   
  
  
     
  
  
Figure 2. An 80 kilometre wide extract from LANDSAT 
Thematic Mapper Band 3 (0.63-0.69um) of Path 125, Row 61, 
16th May, 1992. A linear contrast enhancement has been 
applied to the image. Only noise due to 16-line banding is 
visible due to its linear continuity across the image. 
No in-situ data on water quality are available to enable 
quantification of the image radiance in terms of water 
parameters which would permit estimation of the real 
significance to the user of the error inherent in the observed 
image noise. However, Taylor et al (2001) measured suspended 
sediment from the coastal to the seaward edge of the plumes 
imaged during the May dry season, the same time of year as the 
present image. They estimated Suspended Sediment (SS) 
concentrations ranging from approximately 100mg/l. near the 
coast to less than 10mg/l. at the seaward edge of the plumes. A 
further estimate of SS can be made 
    
by applying existing regression equations derived from 
LANDSAT TM experiments, to the image radiance values. 
Thus a logarithmic model developed by Zia (1993) for water in 
the Pearl River Delta, with a correlation coefficient of 0.992 
and a mean relative error of 2.7% (Eq. 1) 
Radiance (TM3) = -0.3663 + 32.385 log SS, (Eq. 1) 
applied to the image data suggests an increase of approximately 
5mg/l. of suspended sediment for an increase of one DN value. 
This approximates the order of magnitude for sediment in the 
study area observed by Taylor et al (op. cit), since DN for TM3, 
of 38 in the high sedimented area near the coast, fall to 26 at the 
seaward edge of the plumes. The difference in terms of 
sediment is thus (38-26)*5mg/l. ie. a range of 60mg/l. across 
the area sampled by Taylor et al. Therefore, in terms of the 
observed vertical coherent noise, an error of +/-10mg/l. SS 
could be expected due to the vertical coherent noise. 
2.2 Method of noise removal 
Before addressing the vertical coherent noise, the noise due to 
16" line banding was removed using one-dimensional high pass 
convolution filters (Crippen et al, 1989). A series of median 
filters was then applied to the image iteratively using a 3*3- 
pixel kernel until no further change in the image values was 
observed. Secondly a 5*5-pixel kernel was applied iteratively 
to the image. The iterations were stopped when the Standard 
Deviation within the kernel of the previous and currently 
iterated image were similar. The result of this is shown in 
Figure 3. 
Figure 3. Result of iterative median filtering on image data 
The traverse across the image data in Figure 4 demonstrates 
that the noise removal process has indeed removed the local 
noise without altering either the position of the real boundaries 
in the image, or the actual magnitude of spectral difference 
between the different water masses. 
2.3. Signal-to-Noise ratio 
The Signal-to-Noise ratio for LANDSAT Thematic Mapper is 
given as 341:1 (Pease, 1991), although over low reflectance 
surfaces such as waterbodies the ratio is lower due to a greatly 
decreased signal. Noise patterns over water are also more 
visible due to the relative homogeneity of water compared with 
land areas. The signal response for areas of homogeneous 
surface is calculated by averaging the pixels within a window 
of a given size and the noise component is estimated by by 
calculating the Standard Deviation of the pixel responses within 
     
the 
Cul 
The 
the 
afte 
afte 
NY 
  
  
At 
no 
no 
re 
is 
ch